Workflow Hub

I built a quarterly “AI Audit” of my Claude history, adapted from The Rundown’s AI Audit prompt and turned into a repeating loop. Each quarter, Claude reads every conversation I’ve had and reports patterns I can’t see myself: where my stated priorities differ from where my attention actually goes, which decisions I keep reopening, and connections across work, money, travel, and personal life. It then rewrites its own instructions for working with me based on what it found. I request my Claude data export, and the run finds the export email in Gmail, downloads it through Claude in Chrome, and unzips it on my machine through the desktop app. I never touch a zip file. The conversations are condensed and split by month, and Sonnet subagents pull structured notes from each month. An Opus subagent combines the monthly notes into a quarterly report under strict evidence rules: every finding cites at least three conversations across at least two months, states the strongest counter-evidence, and labels each claim as OBSERVED or INFERRED. When the evidence is thin, it says “insufficient evidence” instead of padding. Claude writes and maintains a Google Doc with two tabs. Reports contains one report per quarter with footnoted citations. Progress compares the quarters and includes two sections I added to the original prompt: Stop / Start / Continue / Change, which generates concrete behavior changes for the next quarter, and a Cross-domain pass, which identifies patterns spanning work, money, body, and personal life that I would not connect on my own. I push back on findings I disagree with. Those corrections go into the doc and a versioned prompt, now v1.1, so the next run does not repeat the mistake. Claude then takes the latest audit and its stored memory about me and proposes rewritten project instructions. The first rewrite removed stale facts and added rules to check memory, docs, and connected apps before saying it does not have something; move long-running threads into a maintained doc because one tracking chat had grown to 434 messages; and name the number that would settle an open decision because the audit showed I close decisions once a number exists and keep circling the ones without one. On the first non-holiday weekday after each quarter closes, a scheduled task emails me to request the export while I am at my desk. Scheduled checks later that day pick up the export and run the whole pipeline from a written runbook. I backfilled Q1–Q3 2026 to start, and Q3 alone covered 151 conversations and 2,922 messages. Step-by-step: 1. I started from The Rundown’s AI Audit prompt. 2. I requested my Claude data export and automated the process of finding the export email in Gmail, downloading it through Claude in Chrome, and unzipping it through the desktop app on my machine. 3. I condensed the conversations, split them by month, and used Sonnet subagents to extract structured notes from each month. 4. I used an Opus subagent to synthesize the monthly notes into a quarterly report. Each finding must cite at least three conversations across at least two months, include the strongest counter-evidence, and be labeled OBSERVED or INFERRED. Thin evidence is reported as “insufficient evidence.” 5. I had Claude write the results to a Google Doc with Reports and Progress tabs, including footnoted quarterly reports, cross-quarter comparisons, Stop / Start / Continue / Change recommendations, and a Cross-domain pass. 6. I challenged findings I disagreed with and stored those corrections in the doc and in the versioned prompt, now v1.1. 7. I had Claude use the latest audit and its stored memory about me to rewrite its project instructions, including rules about checking connected sources, maintaining long-running threads in a doc, and naming the number that would settle an open decision. 8. I scheduled the export request for the first non-holiday weekday after each quarter closes, then used scheduled checks and a written runbook to run the pipeline. 9. I backfilled Q1–Q3 2026, with Q3 covering 151 conversations and 2,922 messages. 10. To build your own version, add the evidence rules and OBSERVED/INFERRED labels, split the work into inexpensive monthly extraction and one strong synthesis, write the output to a living doc instead of a chat, and make the final step “rewrite your instructions for me.” That turns the audit from a report into a feedback loop.

I built Lumi, an HR assistant in French, with Claude Code in about two hours. It was an experiment: could Claude Code build a working product from a plain brief while being left completely alone? Employees ask questions, and Lumi answers using their company’s own HR documents. Each company has its own space, and HR can see which questions the documents couldn’t answer. Step-by-step: 1. I set up one window with Claude Code on one side and the files appearing on the other. I didn’t write code, but I could always see what was being built. 2. I described the product in plain language. I dictated my brief out loud, typos included: who pays, who uses it, the problem, a demo client, keeping clients separate, the admin view, the language, and the price. 3. I let Claude Code build the product and make every call. Thirty-four minutes later, the first version had 47 files and 3,949 lines. It even invented a fictional client company with a full HR handbook so I could demo it immediately. 4. I put the product online. Claude Code wrote a setup guide for Supabase and Vercel. I followed it one step at a time and sent a screenshot whenever something broke. 5. I improved it with feedback. I told Claude Code that the design looked like it had been created by a junior designer. Eighteen minutes later, it came back with a name, a design system, pricing plans, and everything in French. 6. I checked security, although I had skipped this during the build. When I asked for a full security review eleven weeks later, it took less than six minutes and found three serious problems, including a database function that allowed anyone to read a company’s documents. Claude Code fixed all three. Run `/security-review` before sharing the link and before each new change. The full write-up, including the unedited prompts: https://www.theaithinker.com/p/how-to-build-a-saas-product-with

I built a module that runs in MagicMirror on a Raspberry Pi and displays a real-time to-do list on a television in my office. Claude does an admirable job managing my daily task list, which is stored in `TASKS.md` on my local computer. I wanted an at-a-glance dashboard showing everything I have in progress, sorted by priority and synchronized with that file. I described the goal to Claude Code. Since I already had an old Raspberry Pi 3 and an off-the-shelf widescreen TV in my office, Claude Code suggested running MagicMirror on the Pi and Syncthing on both the Pi and my local computer to keep the file updated. Claude then built the MagicMirror module to read `TASKS.md`. After a few tweaks to remove the default modules and clean up the spacing between items, the display was up and running. I set MagicMirror to start automatically when the Pi boots, so now my to-do list is in front of me all day. Step-by-step: 1. I identified the need for a real-time, at-a-glance view of my prioritized tasks, which were stored in `TASKS.md` on my local computer. 2. I described the goal to Claude Code and told it about the Raspberry Pi 3 and widescreen TV I already had in my office. 3. I followed Claude Code’s suggestion to run MagicMirror on the Pi and Syncthing on both the Pi and my local computer. 4. Claude built a MagicMirror module that reads `TASKS.md`. 5. I removed the default MagicMirror modules and adjusted the spacing between items on the screen. 6. I configured MagicMirror to start automatically when the Raspberry Pi boots.

I built a workflow for creating a living book that exists online and that I can add to, edit, and share whenever I like. I’m an occasional writer who publishes nonfiction essays about observations in life. I realized I could collect the essays into a book, but life doesn’t take an appropriate pause, so the book and the essays never really end. That raised the question: When should I publish it? I created a local Claude instance to access all the files on my laptop. Local is almost always better than a cloud instance, and you can ask Claude to enable remote control so you can access it from your phone as well. I used Opus 5.5 because I couldn’t be bothered experimenting with lower-level models to see whether they could do the task more cheaply. A friend who uses Claude professionally once told me to always use the highest model available and move down if I ran into usage limits. I’m on the monthly Pro plan. I gave Claude the locations of my existing writing on Substack and Medium.com. It downloaded all my essays and organized them into rough sections. I argued with it about the organization and approved the sections once it explained its logic. I then had Claude create a Ulysses folder for all my essays, organized into sections and chapters. I can use it to create a new essay, edit an old one, or reorganize and reorder the book, then tell Claude to update the online version. I could have had it scan the folder regularly and update the site automatically, but that seemed unnecessarily lazy. Claude created a section of a website I already owned where readers can access the living book: https://enwaken.com/read/ During the process, I also had Claude restructure and redesign my overall website several times. It’s amazing at that sort of task. I’m still amazed that I can say something like "I want blue and violet and black, but maximize the contrast between each of the colours" and it will actually pick colors that complement each other. I wish more website designers knew they could do this. Now I have a version of my book that is spectacularly easy to read online and that I can update whenever I like. It’s easier to navigate than an ePub on my Kindle and certainly easier to update. The whole process took two days, half of which was spent arguing about the organization of the chapters. Step-by-step: 1. I set up a local Claude instance with access to the writing files on my laptop and enabled remote control so I could access it from my phone. 2. I used Claude Opus 5.5 through the monthly Pro plan. 3. I gave Claude the locations of my existing essays on Substack and Medium.com and had it download them. 4. I reviewed Claude’s rough organization of the essays into sections, argued about its choices, and approved the structure once it explained its logic. 5. I had Claude create a Ulysses folder organized into sections and chapters so I could add, edit, and reorder essays. 6. I told Claude to update the online book when I changed the Ulysses folder. 7. I had Claude create the book as a section of my existing website and redesign the site as needed. 8. I now maintain a living online book that is easy to read, share, and update.

Mid- and end-of-year self-evaluations used to cost me a week or two of reconstructing a year of work from memory. I rarely recorded or remembered the “small things,” and I couldn’t easily explain how my daily work connected to our goals. People often forget their most impactful work months later, so I wanted continuous capture instead of annual reflection. I also wanted my work goals and development plan grounded in real evidence rather than guesswork. I built an “Achievement Collector” agent as a sub-agent of my Chief of Staff agent. When my Chief of Staff agent recognizes something important, it formats the accomplishment into an Achievement record, asks whether I want to record it, and sends it to the Achievement Collector agent. SharePoint is the system of record, with fields including Details, Business Impact, Key Insights, Related Goal, Significance, and a flag for accomplishments I’m especially proud of. A Power Automate flow lets me log an accomplishment conversationally instead of filling out a form. A Copilot Studio agent sits on top of the list and reads three work-goal documents—mine, my manager’s, and my director’s—so it can map each accomplishment up the chain. Its instructions tell it to prioritize what I’m proudest of, group work into themes, surface business impact, and never invent anything. The result is that my self-evaluation and goal-progress summaries are now queries against real records instead of writing projects. If you want to build one, start with the data-list fields: capturing impact and insight at the moment of the work is what makes everything downstream possible. The combination of my Chief of Staff agent and Achievement Collector has made work a lot more fun for me and turned me into a laser-focused employee. Step-by-step: 1. I created an “Achievement Collector” agent as a sub-agent of my Chief of Staff agent. 2. I configured the Chief of Staff agent to recognize important work, format it into an Achievement record, ask whether I want to save it, and send it to the Achievement Collector. 3. I created a SharePoint list as the system of record, with fields for Details, Business Impact, Key Insights, Related Goal, Significance, and a flag for accomplishments I’m especially proud of. 4. I used Power Automate to let me log accomplishments conversationally instead of filling out a form. 5. I placed a Copilot Studio agent over the SharePoint list and three work-goal documents: mine, my manager’s, and my director’s. 6. I instructed the Copilot Studio agent to map accomplishments up the goal hierarchy, prioritize what I’m proudest of, group work into themes, surface business impact, and never invent anything. 7. I now query the accumulated records when creating self-evaluations and goal-progress summaries instead of reconstructing my work from memory.

I’m 78, and I run a homestead, a home server, and a senior center, with dozens of jobs underway at once. Most jobs get done a piece at a time, while to-do apps only track whether something is “done” or “not done.” So I built Taskmaster: a plain-text ledger inside a Claude Project. I first tried building a large custom app, then scrapped it. One of my own rules is, “Build the ugly version first.” In the field, I speak notes into the Claude app on my phone. Claude saves my exact words, updates each job’s progress, and asks me one question at a time when something is unclear. When I ask, “What’s next?” it considers deadlines, weather, and my energy level, then gives me one to three things to do. Anything tied to a specific time goes on my Google Calendar. Home Assistant then reads the reminder aloud through my hearing aids. Today, it checked the forecast and scheduled my paint job for the one day this week that stays above 50°F overnight. The machine carries the complexity, so I don’t have to. Step-by-step: 1. I keep each ongoing job in a plain-text ledger inside a Claude Project. 2. When I’m in the field, I speak notes into the Claude app on my phone. 3. Claude saves my exact words, updates each job’s progress, and asks one question at a time when something is unclear. 4. When I ask, “What’s next?” Claude weighs deadlines, weather, and my energy, then gives me one to three tasks. 5. I put anything with a specific time on my Google Calendar. 6. Home Assistant reads calendar reminders aloud through my hearing aids. 7. For example, Taskmaster checked the forecast and scheduled my paint job for the one day this week that stays above 50°F overnight.

I built an AI-assisted Lightroom Classic workflow that lets photographers create branded marketing images without leaving Lightroom. The problem I wanted to solve was simple: photographers often finish editing an image, export it, open another design tool to add a logo, watermark, text, or promotional graphic, and then export it again. I wanted that entire branding step to happen inside Lightroom. Using AI-assisted development, I built Brandstamp Pro, a Lightroom Classic plugin that lets me add branding, preview the result, save reusable presets, and export the finished image from the same workflow. The result is a much simpler process: edit photo → add branding → preview → export Instead of: edit photo → export → open another design app → add branding → export again I originally built it for my own real-estate photography workflow, but the same idea works for photographers who need branded social images, booking graphics, promotional images, watermarks, or client-ready marketing photos. The biggest benefit is not that AI adds the branding automatically. It helped me build a tool that removes a repetitive step from an existing photography workflow. Step-by-step: 1. I selected an edited photo in Lightroom Classic. 2. I opened the Brandstamp panel. 3. I added a logo, watermark, text, or other branding. 4. I adjusted the size, position, opacity, color, and placement. 5. I previewed the branded result before exporting. 6. I saved reusable presets for branding I use repeatedly. 7. I exported the finished branded image directly from the same Lightroom workflow.

In my ongoing mission to go paperless, I created a web app that runs on tablets and replaces paper sign-in sheets. The admin console gives HR instant visibility and detailed reporting on employee and guest attendance across all our branches from a single pane of glass. In Phase 2, I’m expanding the app to eliminate even more paperwork by allowing employees and managers to request, approve, and manage PTO. Department calendars show everyone’s PTO balance and minimum coverage requirements. Step-by-step: 1. I created a web app that runs on tablets to replace paper sign-in sheets. 2. I added an admin console so HR can view and report on employee and guest attendance across all branches. 3. In Phase 2, I expanded the app to support employee and manager PTO requests, approvals, and management. 4. I added department calendars showing PTO balances and minimum coverage requirements.

My wife and I used ChatGPT to bring our vision for a new master bedroom to life instead of hiring a designer. We started by sketching the basic layout, which ChatGPT rendered into a floor plan. We identified the wood, the type of cut we wanted, and the finishes. Then we had a detailed conversation about what we wanted the room to look like from each side. For smaller details, we scanned Google Images and Pinterest for inspiration pictures and designs we liked. We used ChatGPT Voice for back-and-forth conversations and final tweaks, re-rendering the images as our ideas developed. We took the final images to our contractor to help him understand the higher-level vision we wanted to achieve. Step-by-step: 1. We sketched the basic layout for the new master bedroom. 2. We used ChatGPT to render the sketch into a floor plan. 3. We identified the wood, the type of cut, and the finishes we wanted. 4. We described what we wanted the room to look like from each side. 5. We scanned Google Images and Pinterest for inspiration pictures and designs. 6. We used ChatGPT Voice for back-and-forth conversations and final tweaks, re-rendering the images as needed. 7. We shared the final images with our contractor to communicate the higher-level vision for the room.

I built a website that turns property photos, architectural renderings, or even just an address into cinematic property videos. I’m a civil engineer and photographer/videographer, not a software developer, so I used ChatGPT, Gemini, and Claude Code to help build the site and work through problems as they came up. Much of the process involved translating the way I would personally edit a real estate video into usable software. For example, kitchen and living room shots should usually stay together and have longer durations, bedrooms should stay near their bathrooms, an opening shot should feel different from a detail shot, and every clip should include some movement so the result does not feel like a generic slideshow. I used AI to turn those editing principles into actual logic for the site. The system analyzes and organizes the images, builds a storyboard, ranks scenes to decide which ones to use, creates prompts for video models, generates the clips, and assembles everything into a finished video. Most of the development has been iterative. I test the site with real properties I’ve shot, watch the results, identify what feels off, and then improve the prompts, scene planning, timing, or video assembly process. AI has been useful for much more than generating the final video. I’ve also used it to write code, troubleshoot errors, plan features, improve the user experience, and think through how the product should work from the perspective of its primary user. Step-by-step: 1. I defined the video-editing principles I use for real estate footage, including grouping related rooms, varying shot durations, distinguishing opening shots from detail shots, and adding movement to every clip. 2. I used ChatGPT, Gemini, and Claude Code to translate those principles into software logic. 3. I built the site to accept property photos, architectural renderings, or an address as input. 4. I had the system analyze and organize the images, build a storyboard, rank scenes, and select the footage to use. 5. I used AI to create prompts for video models and generate the clips. 6. I assembled the generated clips into a finished property video. 7. I tested the site with real properties I had shot, reviewed the results, and iteratively improved the prompts, scene planning, timing, and video assembly. 8. I used AI throughout development to write code, troubleshoot errors, plan features, improve the user experience, and think through the product from the primary user’s perspective.

I spend a good chunk of my week volunteering for our local community group, Byron Bird Buddies, keeping an eye on local and regional bird news. With so much happening across Australia—from avian health updates to local conservation stories—tracking it all by hand takes a lot of time. I recently tried my hand at vibe coding to see if I could build a workflow that automatically gathers interesting regional bird stories into a handy hub that I could share on our Facebook page. It did not quite fly, mostly because our little volunteer group runs without funding, and my prompting skills probably leave a bit to be desired. I am convinced that a simple, free automated workflow like this would be a game changer for grassroots conservation groups and thousands of birders everywhere. If any clever builders in the community have ideas, tips, or would be keen to tackle a project like this, I would love to connect. I also asked my AI buddy, Al, whether an app like this already exists. This was the answer I received: > No, there is no existing single app or consumer platform specifically designed to aggregate, curate, and publish regional bird and conservation news to social media. > > While there is not a pre-made app doing this specific workflow, the reasons you have not seen one come down to how birding technology is structured and the types of tools currently available: > > 1. Data and sighting trackers, such as eBird, Birda, and iNaturalist, track observations, sightings, and checklist data submitted by birders. They do not crawl news sites, press releases, or government databases for written news stories or conservation updates. > 2. Official disease and threat trackers, such as Avian Flu Watch, DAFF, and Wildlife Health Australia, maintain live tracking maps and statistics for specific issues such as avian influenza. They focus on raw data or single-issue reporting rather than general, human-interest, or local conservation news. > 3. Generic news aggregators and RSS readers, such as Feedly and Inoreader, collect RSS feeds from selected websites into one dashboard. They require manual setup for every website or RSS feed, and they do not automatically summarize, filter by region, or publish directly to Facebook in a readable format without paid extensions or third-party automation tools such as Zapier or Make. > > A dedicated app may not exist because scraping news websites or using live search APIs, such as Google News or Bing News, costs money. Without a subscription model, developers may struggle to keep a free news-scraping app running. Region-specific filtering is also difficult: finding bird news specifically for the Northern Rivers and NSW alongside broader Australian stories requires fine-tuned filtering, often using AI or large language model processing, to avoid irrelevant articles. > > In the meantime, a no-cost approach could combine free RSS aggregators and social schedulers. Possible sources include ABC North Coast, NSW DPIRD, local council news, and BirdLife Australia. Google Alerts could also be set up for keyword combinations such as “Byron Bay” bird, “Northern Rivers” conservation, or “NSW” avian flu. > > Another option would be to copy and paste a digest of links or raw text into a free AI assistant such as Claude, ChatGPT, or Gemini once a day or once a week, using a prompt such as: > > “Summarize these articles into a short, friendly 3-bullet-point Facebook post for Byron Bird Buddies readers.” > > The conclusion was that the idea remains relevant for a community of workflow builders. A free-tier automation platform such as Make or n8n, or a Python script, could potentially fill a real gap for volunteer groups like Byron Bird Buddies. > > Written by Gemini AI for Larry > > Larry Larstead > larrylarstead(at)gmail.com Step-by-step: 1. I identified the time-consuming task of manually tracking local, regional, and Australian bird and conservation news for Byron Bird Buddies. 2. I tried using vibe coding to build a workflow that would gather relevant regional stories into a central hub. 3. I planned to share the collected stories through the Byron Bird Buddies Facebook page. 4. After the first attempt did not succeed, I asked my AI assistant whether an existing app already supported this workflow. 5. I reviewed possible alternatives, including RSS aggregators, Google Alerts, manual AI summarization, free-tier automation platforms such as Make and n8n, and Python scripts. 6. I shared the idea with workflow builders in the hope that someone could offer tips or help create a free solution for grassroots conservation groups.

“Honey, do you remember where Ryan’s practice is? Do you remember our Wi-Fi password? Do you know my TSA PRE number? Remember who that painter was that we used last year?” If you have a family with children, you may hear questions like these all the time. In a large family, there always seem to be questions about the house, the kids, or technology that could be answered more easily. The information is usually stored in different places: password files, a Rolodex of business cards, or sheets of paper in junk drawers. I set out to build an app where anyone in my family could use voice to save something for the future or retrieve something they had already stored. I integrated AI to understand the meaning of each voice note. It might create a calendar event, store a password in a vault, or simply remember a phone number. Then anyone else in the family could access the information. The app uses multi-factor authentication for anything particularly secret. It is not intended to be a dedicated password protector; I would build in much more security if that were the goal. Instead, it is an everyday-life app for remembering the little things: How long is the warranty on this appliance? Which exact light bulbs did we use here last time? What was our daughter’s email address that we needed to set up on her phone? Being able to use voice to enter information or retrieve it was the key for me. I think that will save people time. As I started adding entries and validating the idea, my wife came up with the name. After a few weeks, I knew she was completely right, because I now use it about 10 times a day—and I hear that exact phrase every time. Step-by-step: 1. I identified recurring family questions about locations, passwords, identification numbers, contacts, warranties, and household items. 2. I built an app that lets family members use voice to save or retrieve information. 3. I integrated AI to interpret the meaning of each voice note and determine whether to create a calendar event, store a password in a vault, or remember a phone number. 4. I made the stored information accessible to other family members. 5. I added multi-factor authentication for information that is particularly secret, while keeping the app focused on everyday memory rather than dedicated password protection. 6. I added entries and validated the idea through regular use, eventually using the app about 10 times a day.

I was charged with leading my book club’s discussion of *Shrines of Gaiety* by Kate Atkinson. The book has many well-developed characters and murder-mystery plotlines, so I decided to create a “Who Is This Character?” game using simple supplies from my home office: envelopes, printed clues, a stapler, and markers. I asked ChatGPT to help design the game. It came up with eight potential characters, then selected the five most important so the game would still work if some book club members were absent. For each character, it created five clues and a simple rules sheet. Characters were assigned numbers and read in order. Players heard each clue without asking questions, but they could guess after every clue. Many characters were identified, and the game led to a meaningful discussion of each character’s motivations. ChatGPT also provided a discussion question for each character, which made the activity easy for me to run. As a bonus, ChatGPT found a 1920s cocktail recipe for the Mother’s Milk drink mentioned on page 336. It included appropriate warnings, nonalcoholic options, and a sample size suitable for an afternoon book club for senior women. Step-by-step: 1. I chose a “Who Is This Character?” game for my discussion of *Shrines of Gaiety* because of its many characters and murder-mystery plotlines. 2. I gathered envelopes, printed clues, a stapler, and markers from my home office. 3. I asked ChatGPT to suggest eight potential characters and select the five most important in case some members were absent. 4. I asked ChatGPT to create five clues for each selected character, along with rules explaining that the characters would be read in numerical order, with guesses allowed after each clue and no questions during the clues. 5. I used the clues and rules to run the game during the book club discussion. 6. I used ChatGPT’s character-motivation questions to guide the discussion after the game. 7. I also used ChatGPT’s 1920s Mother’s Milk cocktail recipe, including its warnings and nonalcoholic options, for the afternoon gathering.

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 build an office platform for small nonprofits, and every week users ask how to do something. I also keep changing features they have already learned, so written instructions go stale quickly. Screen recordings take an afternoon each and can be out of date by Friday. I use Claude Code in my project to turn each user question into both a software change and a how-to video. The video runs against a throwaway copy of the app filled with a made-up organization, so no real data appears on screen. It performs the task as a person would—clicking, typing, and waiting—while displaying captions one step at a time and moving a yellow pointer to each button before clicking it. There is no voice. I tried narration but dropped it because captions can be paused and read, and they never mispronounce anything. The important part is that the script also tests what it films. After each step, it confirms that the screen did what the caption says and prints PASS or FAIL. If anything fails, there is no video yet; I fix the bug first. In one day, the videos found four real defects that our automated tests had missed, so the video is also the test. `ffmpeg` turns the recording into an MP4 and a poster image. I watch and approve the result, then publish it on our videos page and inside the app’s Help box, next to the instructions for that screen. When the screen changes, I run the same script again and get a new video in minutes. The first videos were too fast. I am not as fast as Claude, and the person watching is trying to learn something new. Now each caption stays on screen long enough to read—about 1.5 seconds plus a third of a second per word—before the next click. The workflow produced 35 how-to videos in five days, most of them one to three minutes long. The checking in step 4 started on day four, after the videos kept finding problems. Every video since then has checked itself. Step-by-step: 1. I take a user’s email question or request and use Claude Code in my project to make the requested software change. 2. In the same session, I have Claude write a short script for a how-to video of the task. 3. The script starts a throwaway copy of the app, fills it with a made-up organization, and performs the task by clicking, typing, and waiting. 4. The script displays captions one step at a time and moves a yellow pointer to each button before clicking it. I use captions instead of narration so viewers can pause and read them without hearing mispronunciations. 5. After each step, the script confirms that the screen did what the caption says and prints PASS or FAIL. If anything fails, I fix the bug before producing the video. 6. I use `ffmpeg` to create an MP4 and a poster image from the recording. 7. I watch and approve the video, then publish it on our videos page and in the app’s Help box next to the relevant screen instructions. 8. When the screen changes, I run the same script again to create an updated video in minutes. 9. I keep each caption on screen for about 1.5 seconds plus a third of a second per word before the next click.

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.

My mother recently moved into a retirement suite, which meant downsizing her possessions. We ended up with dozens of physical items she no longer had space for, and we wanted her children and grandchildren to have an opportunity to sign up for the family items they loved before anything was donated. I needed a way to catalog everything quickly, share the information remotely, and verify that we weren't accidentally giving away items of significant value. I tested a physical-to-digital workflow using Gemini. Instead of putting the items in a pile or typing out endless lists, I drew a numbered grid on a large whiteboard, placed one item in each square, and took a few photos. Gemini identified the objects by grid position, categorized them, estimated their market values, and built a ready-to-use spreadsheet with a signup column. The physical staging took about 30 to 45 minutes, depending on how many items were on that particular board. The entire digital workload—including cataloging, valuing, formatting, and creating the spreadsheet—took under 10 minutes. The process turned an overwhelming chore into something manageable and produced an end product that was easy for everyone to use. Step-by-step: 1. I drew a simple grid on a whiteboard, numbered each box, and placed one item in each square before taking photos. 2. I uploaded the photos to Gemini and prompted it to identify every item by its board name and grid number. 3. I asked Gemini to classify each item into logical groups, such as Glassware, Ceramics, and Woodcraft, and provide quick, realistic resale-market valuations based on visual condition. 4. I asked Gemini to generate a downloadable Excel file formatted as a signup catalog, with columns for Item ID, Grid Number, Description, Category, Estimated Value, and a blank "Claimed / Signed Up By" column for family members to enter their names. 5. I uploaded the file to Google Drive, converted it to Google Sheets, and added a note at the top explaining how tie-breakers would work if two family members wanted the same item.

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’ve been experimenting with agentic coding for quite some time, and I’ve noticed that the tasks I let the agent handle keep getting bigger. For larger tasks and projects, I gradually moved to a standard workflow, which I published at https://github.com/evermeer/CodingAgentOrchestration in part 2 of the tutorial. I use this workflow professionally and in my side projects. My most recent side project is Quadrim, a Four in a row game played in a 4×4×4 cube. I made it open source and included all the prompts for both the code and the artwork: https://github.com/evermeer/Quadrim It took me about eight working days to publish the first version to the app stores. Step-by-step: 1. I start by describing the idea. 2. I brainstorm the idea with the agent. 3. I use `to-spec` to create a specification document for the idea. 4. I use `to-ticket` to create independent, executable tasks. 5. I review the specifications and tickets from the perspectives of a senior software engineer, architect, security officer, and product manager. 6. I implement one ticket at a time. 7. I code-review the generated implementation for each ticket. During the pre-implementation steps, I always add this instruction to my prompt: “Ask clarifying questions until any ambiguities are resolved.” Extra fine-tuning of the idea, specifications, or plan pays off during implementation. When I let the agent build code, I ask it to use TDD, perform a code review at the end, and resolve the major and critical findings.

I run an out-of-home advertising agency in Malta, covering billboards, buses, hotel screens, and radio. Our entire business ran on Monday.com boards for CRM, deals, bookings, supplier bills, and bus-campaign reconciliation. The system was expensive, difficult to change, and full of broken formulas. I’m not a developer, but in about two weeks I worked with Claude to replace it with our own CRM built on Google Sheets and Google Apps Script. The team now uses it every day. Step-by-step: 1. I mapped our real processes first, explaining in plain language how we sell, book, invoice, and pay partners. Claude read our existing Monday.com boards and imported the data. 2. I built the system screen by screen, releasing each part and testing it with live data before moving on. The CRM now includes contacts, companies, deals with order lines, bookings, billing, partners, tasks, and a dashboard. 3. I automated repetitive tasks. When I upload a signed order form, the app reads the lines. Bookings are created automatically from deals, and a Gmail side panel and shared mailbox file client emails under the correct contact or deal. 4. I connected QuickBooks Online. The system pulls invoices and supplier bills, calculates each partner’s revenue share according to its own rules, produces the monthly partner report, and creates the bill. I approve anything that writes to the accounts. 5. I turned a monthly spreadsheet chore into an automatic check for bus campaigns. The app compares what we committed to the client with what the bus operator billed us and flags overbilling or margin problems month by month. The first run found a supplier bill that appeared to be a duplicate. 6. I made the system safe to hand over. Every change is logged and can be undone, with user roles and a shared drive in place. We also wrote a short user guide and test checklist before the team started using it. The main lesson I learned was to treat AI like a capable builder who needs your business rules spelled out. Starting with real data, releasing small pieces, and checking each one before adding the next made the build manageable. The tools I used were Claude, Google Sheets, Google Apps Script, Gmail, and QuickBooks Online.