Share your best AI workflow. We could show it to 2M+ people.

I’m building an automated content engine for a client using n8n and the Claude API. Every week it produces blog articles and social posts for eight brands. The client already has a blog with [number] articles, and we want the output to match that voice. What I’ve tried: [e.g. a tone-of-voice description in the system prompt / a few example articles in the prompt]. What I get is text that’s fine but sounds AI-written and slightly translated. It doesn’t read like a native Swedish copywriter wrote it. What I’m weighing: Is it better to extract a tone-of-voice guide from the blog (rules, word choices, sentence length, do’s and don’ts) or to pass example texts as few-shot? How many examples do I need, and should they be full articles or excerpts? Is a separate editor step that only reviews and adjusts tone worth it? How do you handle one shared base voice across eight brands that each have their own variant?

Sometimes I remember something perfectly badly: not its title, creator, or where I found it, but a sound, a color, a vague cover, a sentence, or a period in my life. I know it exists because I loved it, but I cannot name it. I use AI and my digital traces to recover songs, books, videos, articles, and quotes. Memory is not a catalog; it stores traces. AI can turn those traces into searchable hypotheses. Step-by-step: 1. I dump every surviving clue without waiting for a proper search query. The clues can be as rough as “The title sounded like nu nau tau,” “It was ambient or humming,” “The cover was maybe a green desert?” or “I knew it around 2019.” Another example is: “A book about makers and Disney. White 3D-printed figures, tiny people, and bears on a blue cover.” This is raw memory evidence. 2. I separate certainty from guessing by labeling each clue as certain, fairly sure, possible, or probably distorted. I keep uncertain clues, but do not treat them as facts. A wrong attribution or outdated cover does not invalidate everything else. 3. I reconstruct the original encounter. I ask where I was and whether I Shazamed it, saved it, sent it to someone, screenshotted it, or bought it at a book fair. In one case, I remembered hearing a song in someone’s car and Shazaming it because I could not parse the pronunciation. Remembering that capture action pointed to a retrieval source. 4. I search my own collections first: playlists, liked songs, Shazam history, saved videos, old posts, messages, bookmarks, notes, photos, receipts, and reading lists. The object may already be there under metadata I no longer remember. I start with the smallest reliable fragment. 5. I search for “nu” inside my playlists after remembering “nu nau tau.” That recovered Man O To — Nu. The artist and title had blended into one phrase, and the song was sitting in several playlists, including my Shazam collection. A clue does not need to be complete; it only needs to be distinctive enough inside the right collection. 6. I allow memory to scramble metadata. An artist and title may merge, album art may become song art, and a quotation may survive with the wrong poet attached. I split remembered phrases, swap artist and title, try different word boundaries, consider older releases, and search phonetically using similar syllables, vowel sounds, homophones, alternate spellings, or other languages. I also ask what the cover looked like when I knew it, because artwork changes. 7. I use context and cross-reference traces. A year, trip, relationship, or platform can narrow the search more than a half-correct title. Spotify might preserve the song, Shazam the discovery moment, and an old message the recommendation. Old posts are accidental archives, and dead links may still leave names, dates, or descriptions. If memory fails completely, I inspect behavior from the relevant period, such as repeated listening, watch history, purchases, or library loans. 8. I let AI propose candidates and compare them against several independent clues, including sound, subject, period, cover, and personal context. One matching detail is not confirmation. I show titles, covers, creator names, and short descriptions because recognition may succeed where recall fails. Rejections also help: “No, but the cover was more minimal,” “The title was shorter,” or “It was fiction, not nonfiction.” Each correction sharpens the search. 9. I pay attention to the strangest clue. The blue cover with white figures, bears, and a makers-and-Disney theme led to *Makers* by Cory Doctorow. 10. Once I confirm the object, I save the title, creator, link, why it mattered, and how I recovered it. For example: - Clue: “nu nau tau” - Encounter: heard in someone’s car and Shazamed - Search: Spotify/Shazam collection → “nu” - Result: Man O To — Nu - Distortion: artist and title blended together This becomes a retrieval playbook: which clues survive, where traces live, and how my memory distorts them. The overall workflow is: collect fragments → mark uncertainty → reconstruct the encounter → search personal traces → try phonetic, visual, and contextual clues → compare candidates → recognize and confirm → save the result and retrieval path. I do not need to remember the database fields. I need enough fragments to rebuild the trail. Sometimes the best clue is what I did when I first encountered something. Sometimes it has been sitting in five of my own playlists while I spend years wondering what it was called. 😑😭🎧

I built a workflow in which AI agents build a product under the kind of governance I’d expect from a senior engineering team. After 20 years in software engineering, including 10 years running my own businesses, I learned that writing code isn’t the hard part. Keeping control of it is. So I set up a build loop on GitHub. One thing I’ve kept for myself is decision-making: approving plans, answering product questions, and signing off on what the system has learned. The product it’s building is Cordexa, a tool for document-heavy expert work, such as quantity take-offs from engineering drawings or assessing a tender against a company’s own CVs and case studies. It’s designed to run inside the customer’s own cloud, so their documents never leave it. Every figure is cited to the document it came from or flagged for a person. It never guesses. The problem it solves is complex, but the build loop lets the system work in the background and notify my phone when it needs a decision. That allows me to stay focused on my day job. Step-by-step: 1. I start every module with a plain-English plan and approve it before any code exists. 2. One agent writes the tests from the plan. A different agent writes the code, and it cannot modify the tests. 3. An independent reviewer agent checks every change against its plan. Safe changes merge automatically; anything that changes the product waits for my approval. 4. A watchdog checks every 15 minutes for anything that is stuck, fixes it when possible, or tells me why it cannot. 5. Real test jobs run after every merge and are scored against known answers, so progress is measured rather than guessed. 6. I handle the decisions, answer product questions, and sign off on what the system has learned.

I built a custom web app that uses chance to combine random images and texts into new, unplanned pairings. It recreates a Dadaist technique I use often when making photobooks, which I previously did manually with physical prints. I used Replit to create a digital version of the game. It was easy to connect the app to Project Gutenberg for random texts and the Library of Congress photography archive for images. More importantly, I made it possible for users to work with their own image and text archives. So far, it has produced many new and interesting results, and other artists have used it too. Step-by-step: 1. I identified the manual process I use when making photobooks: combining printed images and texts by chance. 2. I used Replit to build a digital version of this process as a web app. 3. I connected the app to Project Gutenberg to pull random texts. 4. I connected it to the Library of Congress photography archive to pull images. 5. I allowed users to work with their own image and text archives. 6. I used the app to create new pairings and shared it with other artists, who used it as well.

I use Claude Desktop on my Mac, connected to Canva and Metricool, to build and schedule social media posts. I used ChatGPT to help me set everything up because it has more history and access to my company knowledge documents. ChatGPT created instruction files for Claude that provide our company context and guidelines. I pointed Claude to our website, which contains many articles, and connected our Canva and Metricool accounts. We discuss suitable topics to post about. Claude reviews the website for relevant articles or landing pages to link to, then suggests a series of Facebook and Instagram posts. It checks the designs in our Canva account, adjusts them as needed, or asks me for photos and other artwork. Claude then builds the posts in Metricool and saves them as drafts for me to review, following our working rules. I review the drafts and move them to pending. Claude checks that they will publish automatically on the correct date and at the correct time, and it creates links with meaningful UTM parameters. After publishing, I can ask Claude to analyze the results, identify what is working and resonating with my market, and help us test different approaches, content, and topics. Metricool captures the data, and Claude uses it to inform the next round of posts. This saves me hours every week, gives me time to work on other things while Claude handles the process, and produces results that are improving each week. Step-by-step: 1. I used ChatGPT to help configure Claude Desktop and create instruction files containing our company context and guidelines. 2. I connected Claude to our website, Canva account, and Metricool account. 3. I discuss potential topics with Claude, and it finds relevant website articles or landing pages to link to. 4. Claude suggests a series of posts for Facebook and Instagram. 5. Claude reviews our Canva designs, adjusts them when necessary, or asks me for photos and other artwork. 6. Claude builds the posts in Metricool and saves them as drafts for me to review according to our working rules. 7. I review the drafts and move them to pending. 8. Claude checks the publishing date and time and creates links with meaningful UTM parameters. 9. After publication, I ask Claude to analyze the results and identify what is resonating with my market. 10. I use the Metricool data and Claude’s analysis to test different approaches, content, and topics in the next round of posts.

MinecraftEdu first launched nearly 15 years ago, but the classroom world we built with it is still growing. Years of students have contributed to the school and town-building project, creating a shared record of their creativity and learning. Getting it running on a modern Mac meant preserving that history and giving the next group of students a way to add their own chapter. Finding the right version of Java—the software MinecraftEdu needs to run—was central to the process. We used OpenAI’s Codex desktop app, powered by GPT-6, to examine the old setup, research compatibility issues, and test what would work. An older Java version launched the client, but resizing its window caused it to crash. ChatGPT read the crash reports and tested a newer Java 8 version, which gave us a setup that could launch the game and handle window resizing successfully. ChatGPT also repaired the server launcher and created standalone apps, using Apple’s Rosetta compatibility layer to run the older software on Apple Silicon. The result is a usable MinecraftEdu setup on modern hardware with the existing worlds preserved. We can now continue teaching and building in a place that successive classes of students have helped create. Step-by-step: 1. We examined the existing MinecraftEdu setup and researched its compatibility requirements on a modern Mac. 2. We used OpenAI’s Codex desktop app, powered by GPT-6, to investigate the old setup and test possible solutions. 3. We tested an older Java version, which launched the client but caused crashes when the window was resized. 4. We had ChatGPT read the crash reports and test a newer Java 8 version. 5. We used the working Java setup to launch MinecraftEdu and handle window resizing successfully. 6. We had ChatGPT repair the server launcher and create standalone apps. 7. We used Apple’s Rosetta compatibility layer to run the older software on Apple Silicon. 8. We preserved the existing worlds so future classes can continue building and learning in them.

I use Instinct, an AI assistant I text or send voice notes to over iMessage, to manage personal to-dos that used to live in my head or in a pile of iPhone alarms. Many tasks come up when I’m away from my desk, so I used to forget them. Alarms didn’t help much either: they went off once, I swiped them away, and the task was gone. Now, when something comes up, I send Instinct a quick voice note. I don’t need to open another app or type. For example: “Remind me tomorrow at 10 AM to respond to book my next dental appointment, and follow up until I say it's done.” Instinct turns the voice note into a clean reminder with a time attached. If I don’t reply “done,” it checks back in and keeps following up until I close out the task. I use the same thread for recurring tasks, partner follow-ups, ad hoc tasks, and advertising to-dos, so everything I owe people lives in one place. Capturing a task takes a few seconds, which means I actually do it. The built-in follow-up keeps things from slipping just because I was busy when the reminder went off, and the workflow lives in iMessage, where I already am throughout the day. Try sending one voice note today with a real task and end it with “follow up until I say it's done.” Step-by-step: 1. When a task comes up, send Instinct a voice note in iMessage. 2. Include the task, when to be reminded, and whether Instinct should follow up until you confirm it is done. 3. Let Instinct turn the voice note into a scheduled reminder. 4. Reply “done” when you complete the task. 5. If you do not reply “done,” let Instinct check back in and continue following up until you close it out. 6. Keep recurring tasks, partner follow-ups, ad hoc tasks, and advertising to-dos in the same iMessage thread.

I used Claude to review the book *Why Machines Learn*. Then I went through each chapter while Claude questioned me at relevant points. To make the concepts more concrete, I used interactive graphs with x and y axes. The first graph shows the relationship between x and a function, along with the slope of the curve. It uses f(x) = x³ with a tangent line illustrating the derivative. I can drag a slider to move the point and watch the slope change, then switch to 3x to compare it with a straight line. The graph demonstrates several ideas: - Sliding x to 0 on x³ makes the tangent flat and the slope 0, so a small nudge barely changes the output. This is similar to a flat spot at the ends of a sigmoid, where learning stalls. - Sliding to 2 and then to −2 shows a slope of 12 at both points. Because the derivative is 3x², it is positive on both sides and the curve is always climbing. - Switching to 3x makes the tangent lie directly on top of the line, with a constant slope of 3 everywhere. - The nudge-check box tracks the derivative closely, showing that the derivative is essentially “change in output ÷ small change in input.” The second graph shows the input and output layers of a small pattern detector. It represents a 2 × 2 image flattened into four inputs, a hidden layer of two neurons, and two outputs. The weights are set by hand so the network can distinguish a vertical line from a horizontal one. I can click the pixels to change the image and watch the numbers flow through the network. The setup uses weight +1, weight −1, ReLU activation defined as max(0, z), softmax output, and all biases set to 0. For example: - Flatten: x = [1, 0, 1, 0] - h1 = ReLU((1)(1) + (-1)(0) + (1)(1) + (-1)(0)) = ReLU(2) = 2 — left column versus right column - h2 = ReLU((1)(1) + (1)(0) + (-1)(1) + (-1)(0)) = ReLU(0) = 0 — top row versus bottom row - Output: scores (2, 0) → softmax → Vertical 88%, Horizontal 12% The graph includes examples for a left vertical line, top horizontal line, right vertical line, and diagonal. Seeing the numbers change as the factors change really helps bring these concepts to life for me. Step-by-step: 1. I used Claude to review *Why Machines Learn* and question me as I worked through each chapter. 2. I used an interactive x-and-y graph to explore how a function changes and how its slope relates to the derivative. 3. I moved the slider across f(x) = x³, including 0, 2, and −2, to observe how the tangent and slope changed. 4. I switched the graph to 3x and used the nudge-check box to compare a constant derivative with the changing derivative of x³. 5. I explored a second interactive diagram representing a 2 × 2 image, four inputs, two hidden neurons, and two outputs. 6. I clicked different pixels and watched the ReLU activations, output scores, and softmax predictions change for vertical, horizontal, and diagonal patterns.

During a period when my medical symptoms suddenly became much worse, the hardest part was not only the symptoms themselves. Pain, dizziness, visual disturbances, brain fog, and acute stress reduced my ability to explain what was happening, make decisions, and keep track of multiple healthcare contacts. That created an awkward problem: the moment I most needed to communicate clearly was often the moment I had the least cognitive capacity to do so. I used ChatGPT as a temporary coordination and communication layer. I kept a symptom diary so I did not have to reconstruct everything from memory. When my symptoms changed, I could compare the new episode with previous days and prepare a concise description. I also used ChatGPT to think through where to seek care. Depending on the severity and type of symptoms, the relevant question might be whether something belonged with my regular GP, an out-of-hours GP service, emergency care, or one of the specialists already involved. Before speaking with a clinician, I asked ChatGPT to help reduce a messy description to a few medically useful points: what changed, when it started, how severe it was, what made it better or worse, which associated symptoms were present, and what had already been investigated. One of the most useful workarounds was extremely simple: if speaking became difficult, I prepared a short note on my phone and showed it to the clinician. At the top, I explained why I was using written communication—for example, that acute stress, autism, pain, dizziness, or brain fog were making speech harder. Underneath, I included my symptoms, timeline, medications, relevant history, and the questions I wanted answered. That mattered because someone arriving alone, visibly anxious, or struggling to speak can easily have difficulty presenting a coherent medical history. The same is true when symptoms reduce concentration, working memory, or language ability. AI did not diagnose the problem. It helped preserve my ability to participate in my own care when some of the abilities normally required to do that were temporarily impaired. During that period, clinicians investigated several possible causes, treated the problems they found, provided medication when needed, and ruled out more serious possibilities. The valuable part of the workflow was not AI replacing a doctor. It was AI helping me reach doctors with a clearer history, better questions, and a communication method I could still use when my normal one was failing. Step-by-step: 1. I kept a symptom diary so I could compare new episodes with previous days instead of relying on memory. 2. I used ChatGPT to help me think through whether to contact my regular GP, an out-of-hours GP service, emergency care, or a specialist already involved, based on the severity and type of symptoms. 3. Before speaking with a clinician, I used ChatGPT to organize what had changed, when it started, how severe it was, what affected it, which associated symptoms were present, and what had already been investigated. 4. When speaking became difficult, I prepared a short note on my phone and showed it to the clinician. 5. I put the reason for using written communication at the top of the note, followed by my symptoms, timeline, medications, relevant history, and questions. 6. I used the clearer history, better questions, and written communication method to remain involved in my care while clinicians investigated possible causes, treated problems they found, provided medication when needed, and ruled out more serious possibilities.

Most AI agents start with someone writing a prompt from scratch. I wanted a better way. So I built an Agent Creator. I describe the kind of agent I need, and it determines whether I actually need a new one, figures out how that agent should work, creates it, and adds it to the rest of my agent team. Step-by-step: 1. I describe what I need by telling the Agent Creator what I want the new agent to do. 2. It checks what already exists by reviewing my existing agents and skills. If something already does most of the job, it recommends improving or reusing that instead of creating another overlapping agent. 3. If a new skill is needed, it researches the role, including current best practices, common mistakes, useful tools, and what good work looks like in that area. 4. It creates the agent by defining its job, required information, outputs, available tools, and the steps it should follow. 5. It gives the agent the right skills by creating or reusing supporting skills, including examples, reference material, and checks that help it work consistently. 6. It sets clear boundaries so the agent knows what it should handle, what it should not handle, and when another agent should take over. 7. If the new agent belongs in an existing workflow, it adds the agent to the team by updating the handoffs so the other agents know when to use it. 8. Before finishing, it checks the agent’s work by running validation checks to confirm that the new agent follows the standards I’ve set for the whole team. The result is that I don’t have to manually design every new agent from scratch. I can describe the kind of help I need, and one agent can research the role, create the new specialist, connect it to the rest of the system, and make sure it’s ready to use. In other words, I built an AI agent that can help grow its own team.

AI is useful for researching investments, but most workflows stop at “What should I buy?” The harder part comes afterward: Does the idea make sense given what I already own? What would make me add to the position? When should I take profits? What evidence would prove the original thesis wrong? And how do I track all of that without constantly watching the market? I use AI to turn a one-time investment research session into an ongoing decision and monitoring system. I start by asking AI to research the market for potential opportunities. In my case, I look specifically for strong mean-reversion trades, but the same workflow could start with value opportunities, macro themes, sector rotations, individual stocks, crypto, or almost any other investment strategy. Then I give AI my actual portfolio so it can evaluate those ideas in context. After I decide which recommendations I agree with and manually make the trades, I have AI convert each investment thesis into explicit rules for what should happen next. Finally, I turn those rules into automated monitors that periodically check market conditions and alert me only when something happens that warrants another decision. Step-by-step: 1. I define what I’m looking for by asking AI to research potential investment opportunities using criteria I care about, such as mean reversion, valuation, momentum, macro conditions, risk/reward, or another strategy. 2. I have AI investigate current market conditions and rank the opportunities, narrowing a large universe down to a manageable set of ideas worth examining further. 3. I pressure-test each thesis by asking why the opportunity exists, what could drive the expected outcome, what the major risks are, and—most importantly—what evidence would invalidate the thesis. 4. I provide my current holdings so AI can identify overlapping exposures, concentration risks, hedges, or positions that conflict with the new ideas. 5. I ask AI which existing positions the research suggests reviewing and where new exposure might make sense. The goal is a small number of actionable decisions rather than a giant list of interesting trades. 6. I review the analysis and independently decide whether to buy, sell, hold, or do nothing. I keep actual trade execution under human control. 7. Before the market moves, I define the next decision for every position by asking AI to identify conditions that would warrant reviewing whether to: - Add - Take profits - Reduce exposure - Exit - Reconsider the original thesis 8. I turn those conditions into automated monitors. I have ChatGPT periodically check the relevant prices, yields, economic indicators, news, or other variables. Instead of sending routine updates, I tell it to alert me only when a predefined trigger occurs. 9. When a trigger fires, I return to the original thesis with the new information and decide what—if anything—should change. Instead of using AI for isolated investment recommendations, I now have a repeatable loop for managing an investment thesis over time: Find an opportunity → Understand it → Compare it to what I own → Make a decision → Define what would change my mind → Let AI watch for it The most useful part may actually come after the investment decision. By deciding in advance what evidence would make me add, take profits, or reconsider the thesis, I don’t have to start my analysis from scratch every time the market moves. AI becomes a persistent research and monitoring layer while I remain responsible for every investment decision and trade.

I use ChatGPT to help me prepare for appointments with the healthcare professionals who already follow me, especially when several conditions or symptom patterns overlap. I have recurring neurological symptoms as well as diabetes, and I am followed by different professionals. The difficult part is often not noticing that something has changed. It is deciding which changes are relevant, which clinician is the right person to tell, and what information will be useful during a short appointment. Before an appointment, I give ChatGPT the questions and observations I have collected. I ask which details are likely to matter most to that type of specialist, which questions belong with my neurologist, GP, endocrinology team, or another professional, whether I am missing information that would make the situation easier to assess, and which observations are background information versus points to mention early in the consultation. The useful part is prioritization. I might arrive with ten observations, and ChatGPT helps me reduce them to the few that are most relevant. Occasionally, it promotes a detail from “somewhere in my notes” to “mention this near the beginning because it may deserve investigation.” When an appointment is still weeks away, I sometimes use ChatGPT to design a small, temporary tracking routine. Instead of logging everything, I might record three specific parameters in the morning and evening, or note the time, duration, circumstances, and possible trigger whenever a particular symptom occurs. This gives me a compact record to bring to the clinician instead of relying on memory. The division of roles stays clear: I provide the lived experience and observations, ChatGPT helps organize and translate them, and the healthcare professional interprets their medical significance and decides what should be investigated or treated. As a result, appointments become less about reconstructing several weeks of symptoms from memory and more about giving the clinician a concise, usable picture of what has actually been happening. Step-by-step: 1. I collect my questions and observations before an appointment. 2. I give them to ChatGPT and ask it to identify the details most relevant to the type of specialist I will see. 3. I ask which questions belong with my neurologist, GP, endocrinology team, or another healthcare professional. 4. I ask whether any information is missing and which observations should be mentioned early in the consultation. 5. I reduce a longer list of observations to the points most relevant to the appointment. 6. If the appointment is still weeks away, I use ChatGPT to design a small temporary tracking routine, such as recording three parameters in the morning and evening or noting a symptom’s time, duration, circumstances, and possible trigger. 7. I bring the compact record to the clinician, who interprets the medical significance and decides what should be investigated or treated.

For years, I kept lists of things I wanted to look up someday: names, historical references, places, traditions, techniques, unfamiliar concepts, and strange terms I encountered along the way. Most of them lived in Google Keep. They were little bookmarks for my attention: > “I don’t know what this is yet, but something about it caught me.” The problem was returning to them. Looking everything up separately takes effort, especially when I cannot remember why I saved a particular term. So I use AI to turn a batch of those notes into small doors I can open. The basic idea I give AI a list of terms and ask for short, accessible introductions—just enough to understand what each thing is and decide whether I want to learn more. Then I react. Some entries get a quiet “interesting.” Others get: > “OOOH. Tell me more about THIS one.” That reaction determines the next branch. Step-by-step: 1. I copy a manageable batch from my notes. The items do not need to belong to the same subject. A batch might contain an artist, a religious figure, an architectural technique, a place, and an unfamiliar philosophical term. 2. I keep any context I recorded, such as where I encountered the item, what caught my attention, or whether there was a picture, quotation, object, or question attached. That context can help AI interpret an ambiguous name and reconnect me with the original curiosity. 3. I ask for small introductions. My request is roughly: > “Give me a short, accessible introduction to each item. Explain what it is, what makes it interesting, and a few directions I could explore. Flag ambiguous names or uncertain identifications. Keep the first pass light so I can choose what pulls me.” The first pass should make choosing easy. A deep report on every entry would bury the interesting parts under another backlog. 4. I read the introductions and respond naturally: “More about this one,” “How did that tradition begin?”, “Show me what it looks like,” “Why does that material behave like that?”, or “Wait, is this connected to the thing we discussed earlier?” I do not have to justify why one item suddenly becomes fascinating. The other entries can stay for later. 5. Once something catches, I follow the question that appears. An explanation introduces a new term, an image reveals a detail, the detail raises a question, and the question leads somewhere unexpected. For example, a visit to a small chapel prompted questions about a Madonna statue represented in the style associated with Lourdes. That opened into saints, prophets, religious imagery, and how particular figures become visually recognizable. The chapel provided the first door; the next questions appeared while I explored. 6. I use AI suggestions as invitations. AI can suggest entries I might enjoy based on interests that have emerged in our conversations, helping me notice an unfamiliar item I might otherwise skip. My response still steers the exploration. I can follow the suggestion, ignore it, or become absorbed in something completely different. The useful feedback is often immediate: “Yes, more of that,” “Too technical,” “I want the historical story,” or “The visual part is what interests me.” 7. I keep only what I want to return to. Sometimes I save a useful explanation, source, image reference, or new question. Some discoveries later feed into my creative research database. Sometimes the conversation itself is enough. There is no requirement to turn the discovery into a project, course, post, or finished piece. Learning something because it was interesting already counts. If I save a trace, I keep what will help me reconnect later: what it was, what interested me, and where I could continue. 8. I leave room for changing curiosity. An item that feels dull today may become fascinating after another encounter gives it context. Unchosen entries therefore remain available. A backlog is a reservoir, not a queue I must finish. This also makes mixed lists useful: the next interesting connection may come from crossing subjects rather than staying inside one category. The workflow in one line Collect loose curiosities → ask AI for small introductions → notice what pulls → follow emerging questions → save useful traces → return when something feels alive again. Why I use it This workflow makes the first encounter with an unfamiliar subject easy enough that I actually begin. I do not need a polished research question. I can arrive with a strange word from a note written three years ago and say: > “What was this little creature again?” AI helps open the door. My curiosity decides how far we wander. 😅📚

When medical care, psychosocial support, and administrative follow-up are spread across services, emails, documents, and people over many years, there can be plenty of information but very little overview. A new provider sees snapshots, old decisions hide in email threads, and referrals resurface months later. Reconstructing the full story from memory every time requires almost a second brain. So I built a living chronological timeline. I did not want a summary of what my situation was “more or less.” I wanted: `date → event → action taken` For example: `14 March 2026 — Symptoms discussed with service X — Service refers me to Y; appointment requested.` This shows not only what happened, but what followed. My timeline runs from 2009 to the present and is divided into five documents covering meaningful periods. A new document begins when a new phase in life or support begins. Older periods stay stable, while almost all active maintenance happens in the latest document. I reopen historical files mainly when a missing record resurfaces. Step-by-step: 1. I collect existing traces, including emails, appointments, referrals, medical or administrative documents, formal decisions, and earlier notes. A sent email is often only half an event, so I also search for replies and follow-up. 2. I search by trajectory rather than by one name. I use names, spelling variants, organisations, domains, senders, recipients, subject words, and attachment terms. 3. I read the full conversation. A search result is not yet a timeline item, so I establish what happened, when it happened, who was involved, and what action or response followed. 4. I separate evidence levels. I distinguish between my own report at the time, a response or confirmation, an official decision or document, and an event that seems real but still lacks confirming evidence. I do not replace a missing document with an assumption; the absence itself becomes part of the record. 5. I track people and organisations separately. “Contact with organisation X” can hide that different people suggested, revisited, or formally communicated different things. 6. Before adding a new event, I compare it with the existing timeline. It may add information by filling a gap, replace an incomplete entry, clarify an existing event, or mark a missing record while preserving uncertainty. This prevents duplicate versions of the same event. 7. For substantial historical work, I first create a temporary review document using `date — event — action taken`. It can contain open issues, missing documents, and uncertainties. Only after review do I update the main timeline. For ordinary new events, I skip this step when the source and date are clear. 8. I make uncertainty visible by grouping research notes into “Clearly found,” “Possible identity confusion,” and “Not found or insufficiently documented.” 9. I maintain the living timeline using the structure `year → month → chronological events`. Older information is preserved, duplicates are merged, uncertainties stay visible, and open trajectories remain open. In practice, the five documents behave like version control: the first four are mostly archive, while the fifth is alive. 10. I reuse the timeline for thematic analysis. Chronology is the factual foundation, but events can also be regrouped to reveal patterns in functioning, work capacity, transport, medical developments, administrative burden, communication needs, and realistic conditions for participation. Two workflows emerged: - Historical reconstruction: collect sources → reconstruct trajectory → determine evidence status → review → update the correct historical period - Ongoing follow-up: new event → verify source → record date and action → add to the active document Most records are organised around institutions. A GP sees one part, a psychologist another, and an agency only its own procedure. Emails are chronological within a mailbox, but not across a life. A personal timeline turns that model around. The person becomes the organising axis, not the institution. That lets administratively separate events appear next to each other in time, revealing patterns such as repeated referrals, waiting times, administrative burden, or support disappearing when the need for it increases. The workflow in one line is: `Collect sources → reconstruct trajectories → connect events to actions → distinguish evidence status → compare → review → integrate → maintain → analyse thematically.` The result is not a perfect record, but a steadily improving map of what actually happened, with mainly the newest edge continuing to grow while the past stays calmly in place.

Most AI agents are effectively static. You write their instructions, use them repeatedly, notice where they struggle, and manually tweak the prompt when something goes wrong. Valuable feedback from real work is often lost, so the same mistakes can keep happening. I created a self-improvement flywheel that uses actual agent performance data to improve agents over time. The system collects two kinds of evidence: - Task scores showing how well each agent performs across different quality dimensions - Run telemetry and review outcomes revealing recurring failures, coordination problems, and cases where actual behavior differs from expectations A scheduled weekly cycle analyzes that evidence, identifies patterns, creates improvement proposals, evaluates whether those proposals are safe and broadly applicable, updates agent instructions when appropriate, and measures whether those changes actually improve performance. The goal is not to let agents rewrite themselves freely. It is to create a controlled learning loop. Step-by-step: 1. Collect performance data while agents work. Score important outputs across consistent quality dimensions, and record useful execution telemetry such as failures, decisions, reviewer outcomes, and unexpected behavior. 2. Analyze performance trends on a recurring schedule. Calculate per-agent averages, identify weak dimensions, compare agents, and look for improvement or decline over time. 3. Mine run history for recurring patterns across multiple sessions, including agents that repeatedly struggle, low-quality runs, and cases where expected behavior differs from what actually happened. 4. Turn repeated problems into improvement proposals. Before proposing a change, inspect the agent’s current instructions so you do not add a rule that already exists. 5. Evaluate each proposal before applying it. Check whether the lesson is broadly useful, redundant with existing instructions, or in conflict with established behavior. 6. Separate low-risk and high-risk changes. Automatically apply additive or clarifying improvements. Escalate conflicting changes for human review instead of allowing the system to fundamentally change an agent’s behavior on its own. 7. Look for system-level problems. Analyze patterns across agents to identify quality gaps, missing capabilities, or coordination failures that cannot be fixed by changing one agent alone. 8. Apply approved improvements and preserve the history. Update the relevant agent instructions, archive the processed proposals, and version the changes so they remain inspectable and reversible. 9. Measure whether each change actually helped by comparing agent performance before and after the refinement. If quality does not improve, do not automatically assume the change was useful. 10. Repeat the cycle. As agents complete more real work, the system gathers more evidence and gets another opportunity to improve. Instead of treating agent instructions as static prompts, I turned them into a continuously improving system: Work → Evaluate → Find Patterns → Propose Changes → Refine → Measure → Repeat The important part is that the loop is evidence-driven and controlled. Agents improve from real usage, but low-confidence or behavior-changing updates still require judgment rather than being applied automatically.

I used Astra to redesign my personal wardrobe. I want to dress better, but my absolute worst nightmare is having a closet full of clothes that creates more clutter in my brain. I already have enough to think about, so I built an app that decides what I should wear based on the season and local weather using GPT-6 Astra. I gave it a few full-body photos, my height, weight, sizes, and niche details such as having broad shoulders and needing to tailor my waist. I also shared the colors I like, added a master prompt, and asked it to build a site. The results were incredible. I'm the model in every photo, and Astra's virtual try-on is genuinely impressive. If I can't find a piece online, I update the model and regenerate the pictures. Step-by-step: 1. I opened ChatGPT desktop with GPT-6 Astra and Codex/Sites, then turned on Computer Use and image generation. 2. I started a new chat and uploaded four to eight full-body photos, one clear face photo, and closet photos when available. 3. I pasted the prompt below, filled in my name, city, and sizes, and let it work for one to two hours. 4. I sent a couple of correction passes instead of rebuilding the whole system. I removed things I would never wear and added things I actually wear. PROMPT: GOAL Build a working website called [YOUR NAME]'s Wardrobe. It blends two things: 1) A closed uniform system of exactly 30 looks 2) A personalized shopping portal where I am the model in every photo, Aritzia/Uniqlo catalog quality (seamless studio backdrop, full-body try-on, product-card grid) SITE NAME [YOUR NAME]'s Wardrobe THE ONLY INVENTORY - 7 summer outfits - 7 fall outfits - 7 winter outfits - 7 spring outfits - 1 gym outfit - 1 lounge outfit = 30 looks. Nothing else. Each look is complete: top + bottom or one-piece + shoes + at most 2 extras. Reuse pieces across the 7 looks in a season. Cap unique garments at 35–50 including shoes and outerwear. Throw everything else out. WHO I AM - Name: - Lives: [city] - Height: - Weight: - Sizes (top / bottom / shoe): - Body notes: - Work dress code: - Weekend life: - Style in 5 words: - Colors that work / colors I refuse: - Budget for gap-filling buys: - Hard constraints: Attached photos are the identity lock. Reproduce my real face AND real body in every try-on. Do not slim, lengthen, or beautify me. LIVE WEATHER (required, not a mock) On every page load, fetch live weather for [CITY] from Open-Meteo with no API key. Use the correct lat/long and timezone. Show on the homepage: - “Today in [CITY]” - apparent temperature, condition, rain yes/no - ONE recommended look from the 30 - why that look won - 1 weather swap (if rain starts / if it drops 5°C) Selection logic: - apparent temp ≥ 20°C and dry → Summer pool - 15–19°C dry → Spring pool Mar–May, Fall pool Sep–Nov, otherwise the closer season - 8–14°C → Fall pool, prefer the look that already includes a mid-layer - ≤ 7°C → Winter pool - rain now or daily precipitation ≥ 1mm → jacket + closed shoes, no white sneakers or silk - wind ≥ 25 km/h → prefer a layer - gym / lounge days use those uniforms, add a layer only if ≤ 10°C Never invent an outfit outside the 30. SITE STRUCTURE - Home: today + try-on + wear-this checklist - Summer / Fall / Winter / Spring: 7 look cards each, me wearing the full look - Gym / Lounge - Pieces: every unique garment on me, marked OWNED or BUY - Purge: sell / donate / trash for anything not in the system - Rotation: 4-week calendar per season. Weather can override, but it still has to be one of the 7 (or gym/lounge) VISUAL BAR Premium catalog photography. Gray/white seamless, even light, full body. Do not clone another brand’s logo. This is [YOUR NAME]'s Wardrobe. HOW TO BUILD Step-by-step: 1. Build the 30 looks from my photos, stats, climate, and closet photos. Use owned pieces first. 2. Generate consistent try-ons of me for every look and every piece. 3. Build a real clickable site. Hook live weather. Do not fake it. 4. If an image breaks my face or body, regenerate it before shipping. 5. No payments. Personal wardrobe OS only. OUTPUT Live site, the 30 looks, piece list (owned vs buy), ranked shopping list, purge list, weather mapping, and what you inferred vs what came from my photos. Start now. Make the call if a detail is missing. Only ask if the photos are unusable.

I built a custom news agent that skips the Google News and Apple News walled gardens. Eight prompts turn Claude into a personal news editor that reads my trusted RSS feeds, scores every story with a decision model, and delivers a clean daily page for about a dime a month. Each week, pulling two or three good stories for my newsletter digest meant searching multiple sites by hand—hours of clicking and scrolling for a handful of usable links. RSS feeds already deliver that content; the missing piece was a judgment layer that could separate worthwhile stories from noise without the cost or latency of running everything through a full LLM. I used Claude Cowork, Jev from TypeSafe AI, Vercel AI Gateway, RSS feeds, Python, and SQLite. Jev is a decision model rather than a chatbot: it returns a probability and confidence score instead of prose, does not hallucinate facts, and scores a story in a fraction of a second. That makes an automated daily judgment call inexpensive enough to run continuously. Step-by-step: 1. I set up a Claude project to hold the scripts, taste guide, database, and page. 2. I defined the goal and sections, then had Claude interview me to create a short “taste guide” describing what counts as a good story. 3. I found and live-tested the RSS feed for each trusted source instead of guessing feed URLs. 4. I fetched all working feeds into a deduplicated SQLite database and logged failures. 5. I used Jev through Vercel AI Gateway to score every story: first filtering ads and promotions, then ranking stories by section against my taste guide. 6. I selected the winners for each section, with Jev identifying near-duplicate stories so they would not appear twice. 7. I generated one-line summaries and fact-checked them against the source so nothing was invented. 8. I rendered a simple HTML page and scheduled the entire workflow to run every morning. A few details matter. I used Cowork rather than a plain chat because several steps require Claude to work with files, run code, and access the network. I also created a Vercel account and AI Gateway API key manually before the scoring step. Finally, I saved everything to the Claude project instead of a local folder, which allows the morning run to happen in the cloud on its own schedule. Scoring costs roughly $0.10 per month with Jev, compared with a few dollars for a full LLM doing the same job. The workflow turns a two-hour weekly slog into a page that is ready every morning, with each link going directly to the original publisher rather than a Google or Apple aggregator. Full build walkthrough: onabetternote.substack.com/p/build-your-own-news-agent-with-claude — the Substack post includes the exact eight prompts to copy and paste, plus more detail on why a model like Jev matters for cost containment as agents become routine.

Build a Self-Improving AI Meeting Protocol for Teams This workflow was inspired by a question Brooke Benson posted in the Rundown Hub about meeting transcripts that miss action items or assign them to the wrong person. That felt familiar because I used to take meeting minutes as a management assistant. I often had to ask: Who said this? Was it a decision or only a suggestion? Who owns this action? Was the deadline confirmed? As a human note-taker, I could ask the people in the room and summarize while listening. A speaker might say: “Maybe Friday, but that depends on Marc.” My notes might become: “Friday is tentative and depends on Marc.” A transcriber captures more, but it also captures the verbal spaghetti. So the problem is not only transcription quality. It is ambiguity. Instead of only asking AI to interpret a messy transcript afterwards, what if the meeting itself produced clearer signals? The system has four layers: Step-by-step: 1. Meeting grammar Speaking habits that clarify speakers, decisions, actions, ownership and uncertainty. 2. Team information Context about names, roles, responsibilities, terminology and decision scope. 3. Output template A fixed structure for the meeting report. 4. Feedback loop AI compares corrections with its previous output and suggests improvements. The meeting grammar should stay lightweight. At the start, each participant states their name and role. Name the next speaker when possible: “Marc, go ahead.” “Sofie, what do you think?” Unclear contributions can be restated: “So Marc, Friday is possible only if the technical review is complete. Correct?” Mark proposals, decisions and actions differently: Proposal: “I suggest we move the launch to Friday.” Decision: “We are moving the launch to Friday.” Action: “Marc will update the schedule by Wednesday.” Tentative action: “Marc may update it, pending confirmation.” At transition points, repeat what was agreed: “The decision is Friday.” “Marc owns the schedule update.” “The deadline is Wednesday.” Uncertainty should also be explicit: “Owner to confirm.” “Deadline is provisional.” “This is a proposal, not a decision.” The AI should receive these rules directly: - Treat explicit decision language as confirmed. - Do not assign an owner unless ownership is clear. - Label tentative deadlines as tentative. - Flag uncertain speakers instead of guessing. - Prefer later explicit summaries over earlier ambiguous wording. The AI should also receive team context: name, role, responsibilities, decision authority, terminology and communication patterns. Example: Pieter Vermeulen Operations Usually responsible for scheduling and logistics. Often uses tentative language for proposals. Only assign ownership when he explicitly confirms it. These notes should stay transparent and editable. People should be able to review, change or reject notes about their own communication patterns. The report should use a fixed structure: Meeting title Date Participants Main topics Decisions Action items Owner Deadline Status Open questions Risks or dependencies Items requiring confirmation TL;DR This can be built into a shared internal assistant. It receives the meeting grammar, team information, transcript and output template. It should be conservative about certainty: Owner unclear → “Owner to confirm” Suggested deadline → “Tentative deadline” Unknown speaker → “Speaker unclear” Discussed proposal → “Proposal discussed, not confirmed” A human still reviews speaker attribution, decisions, owners, deadlines and confirmed items. Then the corrected report is fed back to the AI: “Which errors came from unclear meeting signals? Which came from missing team context? Which came from your interpretation? What should be improved?” The AI can suggest updates to the meeting grammar or team information. Example: Error: Several people discussed a task, but nobody clearly accepted ownership. Suggested grammar update: “When assigning an action, confirm owner and deadline in one sentence.” Or: Observed pattern: “Pieter often says ‘we should’ when proposing something.” Suggested profile update: “Treat ‘we should’ as a proposal unless explicitly confirmed.” Profile changes should be reviewed before they are stored. The workflow: meeting → transcript → AI report → human correction → error analysis → grammar/team-context update → review → next meeting Over time, fewer errors should need correction because the system becomes easier to interpret. The division of labor is simple: The transcriber captures everything. Humans resolve meaning and ambiguity. AI structures, summarizes and detects patterns. AI output quality depends not only on the prompt, but on the signals the AI receives. Sometimes the best improvement is to reduce ambiguity before it reaches the model.

Like many people, I had tasks scattered across emails, meeting notes, reminders, recurring responsibilities, and things I was simply trying to remember. Standard task managers helped me store tasks, but they did not solve the harder problem: turning unstructured information into a reliable daily and weekly execution system. I used ChatGPT Work and Codex to build a private, responsive task management application around the way I actually work. The system combines AI-assisted task capture with a Command Center, daily planning, weekly reviews, task lists, a Kanban board, recurring tasks, deadline reminders, search, filters, subtasks, comments, attachments, and a complete activity history. The most useful part is the connection between AI and execution. Emails and free-text descriptions can be interpreted with the OpenAI API and converted into structured tasks, reducing the amount of manual copying and organizing required. Step-by-step: 1. I mapped my real workflow by identifying where my tasks came from and what information I needed to manage them properly: title, description, status, priority, category, deadline, responsible person, subtasks, comments, attachments, recurrence, and activity history. I deliberately designed the system around my existing working habits rather than adapting my work to a generic task management template. 2. I used iterative conversations with ChatGPT Work and Codex to define the requirements, review the interface, build the application, test it, and refine individual functions. Instead of creating one enormous prompt, I worked in short cycles: describe a problem, implement the change, test it with real data, and improve it. 3. I built the application as a responsive web app that works across computers, tablets, and phones. Access is restricted through authentication, an approved-user allowlist, and server-side authorization because the system contains real personal and professional tasks. 4. I migrated my actual task history rather than starting with an empty demonstration: 238 tasks, 37 categories, 25 subtasks, 5 comments, and 671 activity records. I preserved invalid or disconnected historical records in a separate archive instead of silently deleting them. 5. I connected the application to the OpenAI API. The AI can interpret emails and free-text task descriptions and help turn them into structured, actionable tasks. The application also supports an email-to-task workflow, so actionable emails do not have to remain buried in the inbox. 6. I built a daily Command Center that gives me an overview of what requires attention, including deadlines, priorities, task status, and upcoming work. I use the daily planning view to decide what to focus on rather than simply working through the newest emails. 7. I added two complementary execution views. The task list is useful for searching, sorting, and filtering a larger number of tasks. The Kanban board gives me a visual overview of progress; tasks can be dragged between five status columns, and the new status is saved automatically. The default task list shows the newest tasks first, making newly captured work easy to find. 8. I kept the context inside each task by allowing every task to contain subtasks, comments, attachments, and a complete activity history. This keeps the reasoning, follow-up, and progress connected to a task instead of spreading them across several applications. 9. I automated recurring work and reminders. Recurring tasks are recreated according to their schedule, while deadline reminders help surface tasks before they become overdue. This is particularly useful for responsibilities that are important but easy to forget because they do not arrive as new emails. 10. I run a weekly review to check overdue work, upcoming deadlines, open commitments, and tasks that have stopped moving. I can then reprioritize, update statuses, and prepare the following week from the same system. 11. I preserved portability and control through Excel import and export and a full JSON backup. This gives me control over my information and reduces the risk of becoming dependent on one interface or platform. The result is not an autonomous agent making decisions on my behalf. It is a private execution system where AI handles part of the interpretation and structuring, while I remain responsible for priorities and decisions. It has given me one trusted place for capturing, reviewing, prioritizing, and completing work. The tools I used were: - ChatGPT Work - Codex - OpenAI API

Several months ago, my dad was in and out of the hospital. My two brothers and I were trying to coordinate his doctor appointments, manage his medications and potential interactions, and keep track of all the other details involved in his care. At one point, up to seven different doctors were seeing him in the hospital on any given day. It became important to track every medication he was taking, what each one was for, and information such as his weight and other vital statistics. When he returned home, we also had to make sure someone checked on him and his wife every day, helped him stay on schedule with his medications, and recorded his diet, mood, and weight. We initially used Apple Notes, a shared iCalendar, multiple text threads, and a weekly call between the three of us. The mental load was huge. If we needed to find information from the previous week, we had to scroll through pages of Apple Notes to locate it. We also struggled to keep the rest of the family updated. Before my dad passed away, I started using Claude Code and Codex to build a simple tool that would keep everything organized and searchable. It also displayed trends in areas such as his mood, appetite, and vital signs, and included a calendar showing who was covering which days and times. The tool was still fairly basic when he passed away. Afterward, we encountered the administrative headaches involved in closing out his estate. It was far more complicated than we expected. We thought having a will, power of attorney, and other documents meant we were prepared, but we were wrong. I began integrating those lessons—and the things we learned not to do—into the final product, Eldercare Log: eldercarelog.com. I built the final tool with Claude helping draft a PRD, which I then handed to Codex for the coding work. It took a few weeks of refining the product with Codex. The tool is hosted on Vercel, with Supabase and Stripe on the backend, and includes the security features I built into it. It is the tool I wish had existed when my brothers and I were going through this journey before my dad’s passing. Step-by-step: 1. I coordinated my dad’s doctor appointments, medications, vital statistics, and other care details with my two brothers while he was in and out of the hospital. 2. We tracked his medications, their purposes, his weight, diet, mood, appetite, and other vital signs while he was at home. 3. We coordinated daily visits and coverage using Apple Notes, a shared iCalendar, text threads, and a weekly call. 4. I used Claude Code and Codex to start building a searchable tool that organized his care information and showed simple trends. 5. I added a calendar to track which family member was covering each day and time. 6. After my dad passed away, I incorporated what we learned from handling his estate, including the things we wished we had known earlier. 7. I used Claude to draft a PRD, then gave it to Codex to handle the coding work and refined the product with Codex over several weeks. 8. I built the final tool with Vercel, Supabase, and Stripe on the backend.