I often run several coding agents (Claude, Codex, Grok) in the same project folder, and I lose track of which one did what. I use mainai, a small command-line tool that scans the folder's sessions and tells me which agent to pick up next. It's useful for anyone who switches between multiple coding agents in one project. 1. Install it with pipx: "$ pipx install mainai" 2. Set your Anthropic key, since mainai requires it: "export ANTHROPIC_API_KEY=sk-..." 3. In a terminal, from the project folder, run "$ mainai ." 4. It reports how many sessions it found and recommends one to continue with. On my last run it found 5 sessions and flagged the Claude session as the most recent and the only one with substantial real work. 5. Follow its advice. For that run, I changed into the project folder and ran claude. It saves me from digging through - maybe a month later - each agent's history to figure out where the work actually left off.

I built Anchors Bay PD, a detective game, on my own using AI at every stage. It's for anyone who wants to make or play mysteries with generated evidence, character voices and music. The platform runs on WordPress and WooCommerce at anchorsbaypd.com. 1. I plan each case in a Claude project that I set up as my case builder. It knows the format, the world of Anchors Bay and the rules of a fair mystery. 2. Claude writes the whole case as structured JSON files: briefing, witnesses, statements, emails, actions, phone calls and the truth behind it. 3. I upload those files into Exhibit Lab, a tool I built with Claude. It turns them into finished visual evidence: statements, CCTV stills, phone records and documents. 4. I make scene photos, character portraits and case covers with ChatGPT Images. 5. ElevenLabs gives every character a voice for voicemails, briefings and team voice notes. 6. Live phone calls to witnesses use ElevenLabs conversational agents. These are built and tested, ready to switch on once there is investment to cover the running costs. 7. Suno makes the music for the in-case radio across six shows, with ElevenLabs voices as the presenters. 8. Claude also coded the platform: ranks and promotions, PDF certificates, a forum, referral rewards, pre-orders and release emails. What it produces: a complete mystery case with visual evidence, character voices, an in-case radio and a storefront with ranks, certificates and pre-orders. All of it is built and run by one person.

I sell fine art photography, so mockups are a big part of how I sell. This workflow is for anyone running an e-commerce shop who wants to cut down on the time it takes to create products and mockups for artwork. 1. I created a skill in Codex that lets an agent pull photography from a folder. 2. The agent uploads that artwork to the app I built with Codex, opendesign, and Higgsfield. 3. The agent uses the app to create mockups for each piece of artwork. 4. I review and approve the mockups. 5. The agent creates the full product listing in Shopify, including pricing, variants, and description. This has saved me hours per week on creating products and mockups.

I built an AI phone coordinator for my chiropractic practice so prospective patients get answers and a booking option even when the front desk can't pick up, including after hours. It's for small chiropractic offices that miss new patient calls because staff are busy or closed. 1. A prospective patient calls the office, including after business hours when staff may be unavailable. 2. The call routes to our custom AI voice agent, trained on a chiropractic knowledge base covering services, office hours, pricing, insurance, and appointment procedures. 3. The AI has a natural conversation, identifies the caller's needs, answers common questions, and determines whether the caller wants an appointment. 4. The AI collects and confirms the patient's name and callback number, plus any relevant appointment information. 5. For appointment requests, the AI checks the connected GoHighLevel calendar and books an available slot when authorized and configured. 6. The patient's information and conversation details are saved to the GoHighLevel CRM, where workflow automations update the lead pipeline and start follow-up actions. 7. Appointment confirmations, reminders, and follow-up messages can be sent automatically, subject to patient consent and messaging configuration. 8. I review conversations, leads, and bookings from one dashboard instead of handling every inquiry by hand. The agent's prompt: "You are a friendly, professional AI patient coordinator for a chiropractic practice. Answer incoming calls naturally, identify the caller's needs, and use the approved clinic knowledge base to provide accurate information. Capture and confirm contact details, help schedule appointments using available calendar slots, and escalate clinical questions or urgent concerns to appropriate human or emergency resources. Never invent pricing, availability, insurance coverage, or medical advice." The goal is to give a small practice the capacity of a larger patient acquisition operation without hiring extra front-desk staff to cover every hour.

At our localization agency, I run this workflow every month for each client. It looks at the edits our professional translators made to the initial AI translations, sorts those fixes by the kind of mistake they correct, and turns the results into suggested updates for the AI configuration. The goal is that future AI translations for that client make fewer of the same errors. 1. Each month, an AI agent for each client analyzes what the translators edited in the initial AI translations. 2. The agent categorizes the changes according to the mistake each edit fixed. 3. Based on those categories, the agent suggests updates to the AI configuration so future translations are better. What it saves: our language teams no longer spend hours categorizing every single change by hand. It is also much quicker, and it finds patterns that would otherwise be missed when people do this work manually.

Claude is my lead developer, and my tester works in another time zone. Test plans, bug reports and follow-up questions all used to pass through me, and with the whole platform in my head, I was the bottleneck. Now the AI and the tester work from a shared board, and I step out of the middle. (Disclosure: the board is an admin screen inside my own platform, WebsitePublisher.ai, but the pattern works with any shared space your AI can reach.) 1. Set up a shared board with topics, file attachments and a read marker for each of three participants: me, the tester and the AI. 2. Schedule an hourly AI round. Claude reads new topics, answers open questions, reviews test reports and logs what it did in a persistent task log. 3. Skip meetings. The tester works asynchronously around the hourly rhythm. 4. Notify me only when a topic is addressed to me, so AI and tester traffic stays out of my inbox. 5. Have scheduled routines post their actions to the board so nothing waits unseen. 6. Claude writes a test plan as a markdown file and posts it as a topic. 7. The tester runs the plan and attaches a markdown report. 8. On its next round, Claude reviews the report. If a fix is needed, it prepares the code change for my review, and another test round follows. 9. When the work is done, the topic moves to the archive. After a couple of weeks, I read summaries instead of relaying messages.

I already use ChatGPT to describe image styles, but I kept re-uploading the same reference image whenever I wanted to revisit a successful look. So I asked ChatGPT's Plugin Creator to build a personal plugin, Stijlvis ("Style Fish"), that saves an image's visual identity as a local text file. It is for anyone who repeatedly asks ChatGPT to extract a style from a generated image. 1. Describe the recurring task to Plugin Creator. Explain what you repeatedly ask for, which inputs matter, and what you want back. In my case, that means using a successful generated image and the conversation around it to extract a reusable style prompt. My brief was: "Create a personal plugin that distils the style identity of a successful AI-generated image into a standalone prompt. It should analyse the chosen image alongside any available generation requests, corrections and approval in the conversation. Separate transferable style features from the specific subject, characters and objects. Return a reusable prompt with a placeholder for a new scene, plus a downloadable .txt file I can save locally. Do not generate another image automatically." 2. Give the plugin rules that preserve your judgement. Mine uses my feedback to understand why an image succeeded, separates essential style features from optional choices such as the original palette, and avoids turning a correction to one character into a permanent rule. The output must also work without phrases like "same style as the reference", or I would still need the reference. 3. Create the plugin, then invoke it where needed. Plugin Creator made Stijlvis as a private plugin. I select it in a relevant conversation and ask it to extract the style, instead of repeating the whole analysis brief. 4. Try it on a real image. My first trial used a sunny illustration of a girl, swans and a rainbow. Stijlvis identified painterly texture, luminous light, coloured shadows, layered depth and softly stylised figures, and kept the scene's particular subjects out of the core prompt. Because the original generation conversation was unavailable, this trial used the image itself. 5. Download the recipe and test it elsewhere. The plugin produced a text file with the reusable prompt, style notes and an optional palette block. I save it locally and change the subject for a future generation. The result is a portable style recipe built from a recurring creative conversation. Whether it preserves enough of the style in a new image is the next experiment, since a text description cannot guarantee an identical result.

I use WriteStack to turn each new Substack article into a batch of short notes for the Substack community. It works from my own writing, reformats it with templates from successful notes, and I polish the results before scheduling. This is for anyone who wants to promote their posts without spending hours writing notes from scratch. 1. I publish my new article on Substack first. 2. I open the WriteStack web app and click "Note generation." 3. I choose my published article as the source. WriteStack integrates closely with Substack, so it pulls the article directly. 4. If the article covers a single subject, I give the AI no prompt. If it covers several topics (as in my newsletter), I instruct the note generator to focus on a single topic at a time. 5. I select how many notes to generate. Limits depend on the tier, and I usually go with about 10. 6. I review each generated note, discarding some and adjusting others to my liking, then move on to Schedule. 7. In Schedule, I set up the batch in an editable calendar. I now create notes in less than half the time it used to take. The notes come entirely from my own material. WriteStack reformats what I wrote, and I do the final polish myself rather than publishing generic chatbot slop.

Late timesheets delayed accurate client billing for our advertising and marketing agency, so I built a Claude project that reconstructs my week from the tools I already use. It drafts time entries that I review and approve before anything reaches our system. It's useful for anyone who logs time after the fact and loses track of smaller tasks. 1. Before each run, Claude checks that my Chrome browsing history and Claude conversation export, which have no connector, are current. If one is out of date, it reminds me and asks whether to wait or continue without it. 2. It pulls the time I've already entered in the system so nothing gets posted twice. 3. It searches the emails I sent or replied to, my calendar meetings (using Gemini notes to confirm they happened and how long they ran), files I edited in Google Drive, Slack threads I joined, work-related browsing, and Claude work sessions. 4. It builds a timeline for each day and groups related items into one block. An email, a doc edit, and a Slack thread about the same issue count once. Personal activity is left out. 5. It estimates time for each block, rates each estimate as high, moderate, or low confidence, and assigns a project and task using my past entries as a guide. 6. It combines related time by day, project, and task before rounding up to the nearest quarter hour, so small items aren’t each rounded up separately, and writes a short description for each entry. 7. It flags unexplained gaps during my usual working hours and asks me about them. It never fills them in on its own. 8. It checks my calendar for PTO and applies our rules for days I worked anyway, so I post only the leave I actually took. 9. It produces a daily reconstruction, a ready-to-enter timesheet, and a weekly reconciliation, then saves a draft to the project. 10. I correct or confirm entries, and Claude updates the totals. Once I approve, it saves a final version. Entering and submitting are separate commands, and it runs each only when I say so. The project instructions require Claude never to invent work or pad hours, never to double-count, and to disclose any source it couldn’t search. Instead of trying to remember my week, I now review entries that already have project codes and descriptions, and it also surfaces work I had forgotten.

I use a four-step system built on Grok Bot to get AI-generated research reports with fewer errors. It's for anyone who relies on AI research and needs to trust the numbers and quotes in it. 1. I start with a research bot that has specific instructions. They cover the types of sources it should use, separating known facts from speculative claims, and cross-checking claims. 2. The report goes to a skeptic. The skeptic is meant to be contrarian but not a roadblock. It asks for evidence, looks for logical fallacies, and creates a short, ranked list of problems. 3. The problem list goes back to the researcher, who then passes the report to my fact-checker. 4. The fact-checker double-checks every number and every quote or paraphrase against the actual document it was found in. The result is a report with fewer errors, but also flags on what's verified and what isn't. Far and away the most valuable workflow that I've built. It costs more in terms of usage, but it's been working great.

I'm a software engineer on a team of five, and we review all pull requests. The review comments, and how they were handled, show mistakes my coding agent has made. I use them to improve the instructions in my Agents.md so future work is of better quality. 1. Ask your agent to review the last six months of pull requests with this prompt: "I would like to improve my Agents.md so that future work is of better quality. Create a file AgentsSugestions.md that list possible changes that you sugest. For this review all PR's from the last 6 months. Examine the comments and look if it was followed up. Only add sugestions to AgentsSugestions.md if you think there is a change it will help future work." 2. The agent reads the comments, checks whether each one was followed up, and writes its suggested changes to AgentsSugestions.md. It only adds a suggestion if it judges that the change will help future work. 3. Review AgentsSugestions.md yourself and decide which suggestions are worth using. 4. Tell your agent to apply the suggestions you chose to Agents.md. What this produces is an updated Agents.md that reflects the problems your team already caught in review.

I built a family archive website to host handwritten letters and photos from a relative's 1942/43 stint in basic training. I built everything through prompts, and it could help anyone with scanned family papers who wants them searchable and open to AI chat. 1. I designed the basic website in Google AI Studio. 2. I migrated the project to Replit for the build-out and hosting. 3. I used Firebase for PDF and photo storage, and a Firebase database for users and data indexing. 4. I built a PDF upload workflow that extracts text from the handwritten letters. 5. I chunk the PDFs for faster retrieval, since the files are large. 6. The workflow creates a search index and a vector database for AI chat. 7. I created server APIs to retrieve data and content, which avoids CORS issues. The result is a site where the letters and photos can be searched and queried through AI chat.

I built Elo, a voice guide on my curb painting site, for visitors who would rather talk than scroll. It's for anyone who wants to see my work, pricing, and a design preview without clicking around. 1. A visitor opens the site and taps "Talk to Elo". The browser opens a WebRTC voice connection to OpenAI through my server, which checks origin, session, and daily limits first. 2. They just talk. They can interrupt and speak any language, and Elo replies by voice. 3. The model has seven tools, and the page reacts to each one. show_completed_work puts real photos from my catalog on screen. show_website_section opens the real testimonials, FAQ answer, work, or pricing section. get_website_information reads exact public site copy, so reviews and prices are quoted, not invented. 4. show_business_links shows WhatsApp or Messenger when they want to proceed. show_competitor_pricing runs only if they ask to compare. 5. propose_curb_design runs only on an explicit request, such as "make it Knicks colors, bold 25". It generates a preview with gpt-image-2 and edits the last image for changes. remember_visit_choices stores only explicit preferences, with the visitor's exact words as evidence. 6. The page sends its visible state back to the model, so Elo only says "here it is" after the image has loaded. Limits are 12 voice calls per visitor per day, 40 per day total, and a capped number of preview generations. Transcripts are saved for private review, with no raw audio. Elo gives one to three short sentences per turn, uses only published pricing and site copy, and can't book jobs or send messages. I confirm the design, price, and timing myself, and previews are labeled as concepts. Result: nobody scrolls or clicks. Visitors talk, and the site shows them my work, pricing, and a preview. Note: Am just building it not yet build.

I turned a photo of a hand-drawn kitchen sketch, covered in handwritten measurements, into an editable Excel sheet. It's useful for anyone who needs a rough materials takeoff from a drawing and wants the uncertain parts flagged instead of guessed. 1. Take a photo of the hand-drawn sketch and upload it to ChatGPT. 2. Send this prompt: "Turn this kitchen drawing into an editable Excel sheet with columns for item description, dimensions (mm), material, quantity and notes. Add area formulas and totals by material wherever there's enough information. If any handwriting is unclear or a measurement is missing, mark it with a red question mark. Do not guess." 3. Have it extract every measurement and item into a consistent table, flagging anything unclear or missing with a red question mark. 4. Review the flagged cells against the original drawing and confirm the units. 5. Ask ChatGPT to save the workflow as a skill, so the columns, conversions, formulas and uncertainty rules don't need re-explaining for future drawings. The result was a structured spreadsheet with uncertain values and gaps flagged for review. The drawing didn't contain everything needed for a complete materials calculation, so checking the source and confirming units stayed part of the process. The bigger takeaway for me was turning a one-off request into a repeatable process. My advice: build uncertainty into your instructions. A red question mark is far more valuable than a confident wrong measurement.

I’ve developed an autonomous, AI-powered genealogical research system that combines historical investigation, evidence verification, family-tree reconstruction, and digital storytelling. Powered by GPT-6, scheduled AI workflows, archival research tools, and structured databases, the system continuously investigates family histories, cross-references historical records, identifies ancestral connections, resolves conflicting evidence, and documents discoveries with traceable sources. Research is managed through recurring, coordinated workflows that prioritize unanswered questions, preserve findings between sessions, and distinguish verified facts from working hypotheses. The goal is to build on previous discoveries rather than repeatedly starting from scratch. The platform supports two independent projects, Family Tide and Dare · Nation Family Archive, each with its own research records, evidence, and publishing experience. Using Supabase, GitHub, Cloudflare, and custom web interfaces, verified research can be transformed into interactive family trees, ancestral maps, historical timelines, digital artifact collections, and richly illustrated family narratives. The vision: A continuously evolving genealogical intelligence platform that doesn't simply collect names and dates. It investigates, questions, connects, preserves, and tells the stories of generations.

I volunteer with Songs of Love, a nonprofit that creates free personalized songs for people living with chronic health and emotional challenges. I built a small AI team to handle the coordination so I can spend my time making each song feel personal. It's for volunteers who want to take on more profiles without losing the human touch. Here is how a profile moves through the workflow: 1. A friend or family member submits a profile of the recipient. When it lands in my inbox, Muse uploads it to Genesis. 2. Genesis writes a first lyric draft using a custom Gem that holds Songs of Love's lyric rules and Suno formatting. The Gem requires a style line at the top in this format: [Style: Insert Genres, front-and-center vocals, crystal clear, high definition, sharp presence]. Muse alerts me to review. 3. I always revise the lyrics. AI drafts tend to fall into predictable patterns, and every song needs to feel written for one specific person. 4. When I finish editing, I tell Muse. I add the song style and music prompts, and Muse sends everything to Suno. 5. Genesis creates the Song Sketch, the cover art included with every song, using a second Gem with the foundation's artwork guidelines. 6. When the music, lyrics, and cover art are complete, Muse alerts me for a full review. 7. Muse attaches the .mp3 and .png to the Songs of Love submission form. I do one final check, and Muse sends it. Then the team moves on to the next profile, and I get to listen to the music. The result is a finished, personalized song with cover art for each profile, and my time goes to the creative edits and final checks.

I often remember an idea by what it meant, not by the title of the conversation where it appeared. Archiefvis ("Archive Fish") is a private ChatGPT plugin I built to turn that kind of memory into a repeatable search through my chat archive. It is for anyone whose exported conversations keep growing faster than their filenames make sense. 1. Check the sources first. ChatGPT lists the archive files it can find, with upload or modification dates and export coverage where known. I confirm they are current, since an older ZIP does not become the latest version just because it was found. 2. Search from the remembered meaning. I give it themes, approximate wording, related concepts and a possible period. I search several clues separately, including spelling variants, using indexes and summaries to find candidates and the conversation texts when those are not enough. 3. Verify the original passage. For each candidate I check the surrounding messages and record the title, actual message date, filename and message number. My own words stay separate from ChatGPT's interpretation and from later summaries. A related passage from a year later is a lead, not the original explanation. 4. Collect only the conversations worth reading. ChatGPT prepares a small Python script using the exact filenames from the index. I run it locally. It copies the selection into one ZIP, checks for missing or duplicate files, and leaves the originals intact. 5. Read the selection and trace the idea. I upload the ZIP and ask for a synthesis of recurring concepts, changes, corrections, tensions and connections, with references back to the conversations. I require complete reading in manageable sections, since opening a file or finding a few keywords is not the same as reading it. The starting prompt I reuse: "First show me which archive versions you can access and let me confirm them. Then search for this remembered idea: [description]. Distinguish verified passages, later repetitions and uncertain candidates." In one session, the collection step worked for five selected conversations, which became the basis for a synthesis of my thinking concerning one topic. The plugin needs accessible files or Library tools; it does not grant access to my computer by itself.

I built a set of agents that run every day to find local businesses with no website, build a site for a selected business, and email the mockup to that lead. It's for anyone who sells web services to local businesses and wants to automate the prospecting. 1. Each day, my Gemini and ChatGPT agents check the major cities in the state I select, for the industry I select, and find the best businesses that don't have a website. They add those businesses to a spreadsheet. 2. A second agent reads the spreadsheet, selects the best prospect based on its need for web services and likelihood of buying, and starts building its website. 3. That agent checks the business's social media for images and its Google Business Profile for information. It also looks for a logo. If there isn't one, it creates a logo for them. 4. The agent finishes building the website, which becomes the mockup. 5. A third agent takes the mockup and sends an email with it to the lead. 6. A fourth agent is set up to reply once the lead responds, book them on a call, and put the call on the calendar. The result is a workflow that takes selected businesses from a prospect spreadsheet through a website mockup and follow-up, with replies leading to a booked call.

I wanted an intricate adult coloring page based on a thistle, but I didn't want a generic floral mandala. So I combined an AI-generated symmetrical design with a photo of a real plant. This is for anyone who wants a custom botanical coloring page. 1. Ask ChatGPT to generate a botanical mandala inspired by your chosen plant, viewed from above. Specify radial symmetry, intricate details, crisp black outlines, closed shapes for coloring, and a pure white background. 2. Review the first result and keep the design you like. This becomes your foundation. 3. Upload a photograph of your plant, preferably taken directly from above. I used a thistle rosette growing in the garden, with beautifully irregular, spiky leaves. 4. Use this follow-up prompt: "Keep the existing mandala as the central design. Integrate the plant from my uploaded photograph around it, using its natural leaf shapes, veins, jagged edges, and radial growth pattern as inspiration. Create one harmonious botanical mandala suitable for adult coloring. Use crisp black outlines, closed colorable shapes, a pure white background, and no shading, text, or borders." 5. Check the details, line clarity, and overall composition, and ask for adjustments where necessary. The result is a custom thistle mandala that combines the precision of AI patterns with the imperfect geometry of a real plant. My takeaway: let AI create the structure, then let a photograph add the character. I expect this to work with ferns, leaves, succulents, or even tree roots too. Thanks gpt :)

My wife came up with this workflow, and I recently copied it for my own invoicing. It turns a finished project into an invoice, a bookkeeping entry, and a drafted client email in one pass. It suits anyone who bills several clients and sends records to an accountant. 1. She keeps a log of each client's billing details and email address, along with the details of her own entity. Claude uses this log and data in a project. 2. When a client project is done, she tells Claude the client's name, the work completed, and the amount due. 3. Claude pulls the rest from the log and prepares the invoice in the format she defined, using a sample given at the start. It adds the relevant line items and assigns the next invoice number. 4. Once the invoice is ready, Claude adds a row to the Google Sheet our CA uses. The row includes the date, client, invoice number, amount, and the invoice file. The file is also saved in a separate Google Drive folder, so the books stay current as I work. 5. Once logging is done, Claude drafts a brief standard email in Gmail to the client with the invoice attached. 6. She sits in the final layer, reading everything, checking the invoices, and changing the wording if needed before hitting send. Invoicing, logging it for accounting, and sending emails used to be four small jobs in four places. Now it is one message to Claude and one review. It saves a lot of time and helps make sure invoices always go out on time.