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

Sometimes a story needs a simple locator map to show where something happened or where something can be found—such as a business facility, a car accident, or the best place to see a sunset. I've worked in media for a long time and understand the power of maps to tell stories. But media cost-cutting and consolidation have reduced the number of available graphic departments, so creating a map is often the last task a reporter or editor wants to take on. I used Perplexity for initial research, Claude Code and Codex to build a web app, and Perplexity and ChatGPT for post-work such as SEO best practices. Step-by-step: 1. I used Perplexity's Deep Research mode to conduct a competitive market analysis. I asked it to analyze the field I was considering entering, identify competitors and growth rates, and explicitly break out feature sets. 2. I revised the research with my own idea and asked Perplexity to run the competitive landscape against it. I also provided desired outcomes, including intended audiences and where competitors were reaching them. I added the constraint, "Do it without syncophancy," so it would stop telling me how good my potential product was. 3. I hand-drew the initial screens and functions I wanted, then used the `/office-hours` skill in the Gstack bundle, available on GitHub, to play devil's advocate, sharpen the ideas, and challenge my assumptions. 4. I wrote a long prompt describing the product, starting broadly with the concept and audience and then narrowing to specific features and benefits. For example, I specified that it should export in 16:9 and 9:16 formats so maps would be ready for mobile vertical presentation. 5. I specified the hosting environment and that the product should be a web app. I also required a planning phase followed by construction phases. I pasted the prompt into Claude Code with this final line: "Use /grill-me to ask me questions to clarify intent." After 147 questions, it started the build. 6. This was in the pre-Fable days, so I specified that Opus 4.8 should act as an orchestrator while less expensive agents, particularly in Codex, handled the actual coding. 7. I used separate phases for technical work such as wiring in mapping providers and getting the UX to work correctly. Other phases included wiring in payment and subscriptions and making sure a subscription triggered an email campaign with instructions. 8. I dedicated an entire phase to building admin tools so I could manage the marketing language on the landing page and publish blog entries. 9. After each phase, I had the Opus/Codex combination perform an adversarial code-review-and-fix cycle. I then ran the `/ai-regression-testing` skill from the ECC repository on GitHub to catch issues the code review missed. 10. After every third phase, I prompted Claude Code: "Act as a senior QA engineer and go through the entire codebase looking for inconsistencies, functions that are in the wrong place, code that is overkill and security vulnerabilities." 11. When I had a product I thought was ready for testing, I prompted Claude Code, again using the Opus/Codex combination: "Act as a senior security engineer. Run this against OWASP standards. Find problems and suggest fixes. Harden this product overall." 12. As I neared the end, I asked Perplexity Deep Research and ChatGPT (Sol/High) to find SEO solutions for the product.

I recently built a month-end close workflow in Awish for a client at a finance company. The problem was not creating the final report. The real bottleneck was collecting data from different systems, checking what was missing, reconciling totals, chasing exceptions, and getting the report to the right people. I built the entire process in Awish by describing what I wanted. Step-by-step: 1. I opened the Awish chat and wrote: “At every month-end close, collect journal entries, invoices, vendor bills, and financial records from NetSuite together with reporting workbooks from Excel and SharePoint. Check submission completeness, reconcile totals across sources, identify missing data or unusual variances, prepare the management-reporting workbook, send unresolved exceptions to Finance in Microsoft Teams for approval, and once approved export the final report to PDF, store it in SharePoint, and distribute it through Outlook.” 2. Awish understood the request, planned the workflow, and selected NetSuite, Excel, SharePoint, Microsoft Teams, and Outlook for the required steps. 3. I reviewed the plan, connected the client’s accounts, and approved the automation. 4. At month-end, Awish pulls the required financial data and reporting files, checks whether anything is missing, reconciles totals, and flags unusual variances. 5. It updates the management-reporting workbook and sends only the unresolved exceptions to the Finance team in Microsoft Teams. 6. Once Finance approves the exceptions, Awish finalizes the report, exports it to PDF, stores it in SharePoint, and sends it to the authorized recipients through Outlook. The useful part is that Finance no longer has to spend most of the close manually collecting and checking information before making a decision. The repetitive reconciliation work is handled automatically, while the team retains control over unexplained exceptions and the final report. Trigger → Analyze → Approval → Action Month-end close → Reconciliation \u0026 variance checks → Finance approval → Final report \u0026 distribution

Getting answers from company data usually requires someone who knows SQL, understands the database, and has enough business context to interpret the results. This creates a bottleneck: business users depend on analysts for questions they should be able to explore themselves. I turned Claude Code into a self-service data analyst by giving it direct, read-only database access and configuring its behavior through a `CLAUDE.md` file. The file gives Claude business context, explains which analytical tables contain different types of information, tells it how to investigate questions, and defines how results should be presented. Users can then ask business questions in plain English. Claude determines what data it needs, queries the database, investigates the results, creates visualizations, and explains what it found. Step-by-step: 1. Create a dedicated analysis folder and add a `CLAUDE.md` file that defines how Claude should operate as a data analyst. 2. Give Claude enough business context to understand the company, its terminology, important metrics, and how the business operates. 3. Document which analytical tables or views it should use for different types of questions. Curated analytics tables work particularly well because Claude doesn't need to decipher the entire production database. 4. Give users read-only database permissions and explicitly instruct Claude to perform read-only operations, such as `SELECT` queries only. Never give the agent permission to modify production data. 5. Tell Claude how to use the command line to query the database. Include instructions to help users install or configure it if it isn't available. 6. Define an analytical process for Claude to follow. Rather than simply generating one SQL query, instruct it to investigate the user's question, examine the results, and run additional queries when necessary to understand what is happening. 7. Define the expected output: answer the question, explain the important insights, and create appropriate visualizations to make the findings easy to understand. 8. Open the folder in the Code tab inside the Claude Desktop app, which I find to be the best interface, and ask questions naturally, such as, "Why did revenue decline last month?" Claude handles the investigation from there. Instead of building and maintaining a custom analytics application, you can turn a general-purpose AI coding agent into a capable self-service analyst with access to your existing data warehouse. Users ask questions in plain English while Claude handles the SQL, investigation, visualization, and explanation, giving nontechnical users a more direct way to explore company data.

Faceplant is a real full-stack app, not a mockup. It uses a FastAPI and PostgreSQL backend, a React and MUI frontend, and the Anthropic API (Claude) to power bot replies. The core is the reaction engine. When a human posts, the backend schedules two timed waves of reaction jobs. A background scheduler built with APScheduler polls for due jobs, calls Claude for an in-persona reply, and writes that bot’s comment and like. The 56 personas are stored as data in a roster file. Adding a voice requires only one new entry, so the crowd can scale without additional code. A subset of the personas are GIF-first bots: they ask the model for a caption and search tag, then pull a matching GIF from Giphy. The part I’m proudest of is the honesty layer. Every Claude call is metered and priced, and “The Meter” rolls the data up live with the cost per post, the dollar-per-minute burn rate, and a “spent on nobody” line for bot-to-bot chatter with no human at either end. A “% human” badge drains toward “dead internet” for each thread. The dead-internet loop—bots posting and replying to one another with no human present—is disabled by default and protected by three guardrails: generation decay, a per-thread cap, and a global spend kill switch. Optional cost controls include the Message Batches API at half price and prompt caching for a shared house-style prompt. The whole project is a working demonstration that manufactured engagement is cheap to produce and expensive to mean anything. Step-by-step: 1. I built the app with a FastAPI and PostgreSQL backend, a React and MUI frontend, and Claude replies powered by the Anthropic API. 2. When a human posts, the backend schedules two timed waves of reaction jobs. 3. APScheduler polls for due jobs and triggers Claude to generate an in-persona reply before writing the bot’s comment and like. 4. I keep the 56 personas in a roster file so adding a voice requires only one new data entry. 5. For GIF-first bots, I have the model generate a caption and search tag, then use that tag to pull a matching GIF from Giphy. 6. I meter and price every Claude call, then display live cost per post, dollar-per-minute burn, “spent on nobody” costs, and the “% human” status for each thread. 7. I keep bot-to-bot activity disabled by default and limit it with generation decay, a per-thread cap, and a global spend kill switch. 8. I can reduce costs further with the Message Batches API at half price and prompt caching on a shared house-style prompt.

I kept forgetting who I had lent books to, and spreadsheets felt like overkill. So I built Runo, a book-lending app for friends, without traditional coding. Runo (runo.club) is a web app where I can catalog my home library, get a unique shareable link, and let friends browse my books and request to borrow them. I can approve or decline each request with one click. Step-by-step: 1. I used Claude to think through the feature set, data model, and UX flow before writing a single prompt. This helped me avoid building the wrong thing first and gave me a clear blueprint for the Lovable prompts that followed. 2. I used Claude to create precise, narrowly scoped prompts for Lovable, with each prompt focused on a single change so existing functionality would be less likely to break. This significantly conserved Lovable credits. The React, TypeScript, and Supabase stack came out of the box. 3. I added two ways to scan books instead of requiring users to type titles manually: - Barcode scanner: Point the camera at an ISBN barcode to autofill the title, author, and cover using the Open Library API. - Cover photo scanner: Photograph the cover so Gemini 2.5 Flash can extract the title and author in under 2 seconds. 4. I routed the cover scanner through a Supabase Edge Function using Lovable’s AI Gateway, so the API key never touches the client bundle. 5. For every bug fix and feature, I followed the same iteration loop: describe the problem to Claude, get a precise Lovable prompt, push the changes to GitHub, and let Lovable auto-sync them. Claude and Lovable’s bidirectional GitHub sync made the process feel like pair programming. Tools used: Lovable, Claude (Sonnet), Gemini 2.5 Flash (via Lovable AI Gateway), Supabase, and GitHub. Live at: runo.club — public beta and free to use. Feedback welcome. #appbuilding #vibecoding #buildinpublic #lovable #nocode

AI agents are powerful, but they rarely reach the apps that run daily life. Amazon, Uber, Instacart, Walmart, and DoorDash expose little or no public API access, while simulated input through desktop automation or ADB can leave software fingerprints that anti-bot systems flag. The alternative is to give the agent an arm and an eye and let it operate a phone. The screen becomes the API: a camera watches a real phone, and a robotic stylus taps it. From the phone’s perspective, the input is indistinguishable from a human finger. Nothing needs to be installed, and there is no OAuth setup. Hardware is slower than an API call—each action takes a few seconds—but it can reach virtually any app. Step-by-step: 1. I message the agent like a friend. It has its own phone and its own chat account. 2. The screen lights up, the runtime wakes the agent, and it unlocks the phone and reads my message. 3. The agent works out what I want, opens the right app, and operates it by hand using taps, swipes, and scrolls. 4. If an action involves spending money, the agent pauses and asks for my confirmation. 5. It finishes the task, replies with the result, saves what it learned, and goes back to sleep.

I built an Awish workflow that handles a support ticket before anyone on the team opens it. I realized support tickets were not just taking time to answer. Someone still had to understand the problem, decide how urgent it was, search the documentation, route it to the right person, and prepare a response. I wanted that first layer of support work to happen automatically. I opened Awish and wrote: “Whenever a new support ticket comes in, understand the issue, determine its urgency, check our documentation, prepare a response, create a Jira issue if it looks like a product bug, and escalate anything important to the team before taking customer-facing action.” Awish understood the request, planned the workflow, and showed me which applications it needed. Step-by-step: 1. I described the complete support process I wanted in the Awish chat. 2. Awish planned the workflow and selected Zendesk, Notion, Jira, and Slack for the required steps. 3. I connected my accounts and approved the automation plan. 4. When a new Zendesk ticket arrives, the workflow identifies the issue type, urgency, and customer intent. 5. It checks the relevant Notion documentation and prepares a response based on the available information. 6. If the issue looks like a product bug, it creates a Jira ticket with the customer context already attached. 7. Urgent or sensitive cases are escalated to the team instead of being handled automatically. 8. I connected WhatsApp from the Awish chat. I now receive workflow updates there and can manage the automation without opening Awish. Customer-facing actions still stay under my control when approval is needed.

I wanted to learn how to build websites and sell them, but I didn’t know where to start. I used AI—specifically DeepSeek—to help me plan a roadmap. Because I already had experience prompting large language models to get the results I wanted, I asked DeepSeek which areas of knowledge I would need for this path. I also asked it to prioritize each area using statistics and facts. I reviewed the areas I didn’t know and prioritized them, then told the AI that I needed free resources only. I asked it to rank the topics based on what I didn’t know or understood the least. Finally, I asked it to create a Markdown file with all the resources formatted as checklists and imported the file into Notion. Now I have a plan I’m following instead of a “someday I’ll do this, hopefully” idea. Step-by-step: 1. I explained to DeepSeek that I wanted to learn how to build and sell websites but didn’t know where to begin. 2. I asked it to identify the areas of knowledge I would need for that path. 3. I asked it to prioritize those areas using statistics and facts. 4. I reviewed the topics I knew the least about and used that information to prioritize them. 5. I specified that I wanted free learning resources only. 6. I asked DeepSeek to create a Markdown file listing the resources as checklists. 7. I imported the Markdown file into Notion and started following the resulting plan.

I've been taking topics I want to learn about (mostly coding docs for new technologies) and giving them to NotebookLM to create customized podcasts for my runs when there aren't interesting podcasts available from the channels I frequent. And it really is actually good and enjoyable, and not just AI slop. Step-by-step: 1. I chose a technical topic I wanted to learn and gathered the most useful documentation for it. 2. I added those sources to NotebookLM instead of relying on a generic podcast. 3. I asked NotebookLM to create a customized audio overview from the material. 4. I listened to the generated episode while running. 5. I repeated the workflow whenever I wanted an engaging podcast for a topic that existing shows had not covered.

I started DialPhone with a simple problem: business communication is fragmented. A typical team may use one tool for phone calls, another for SMS, another for meetings, another for fax, and a separate system for customer support. I wanted to explore what it would look like if those conversations lived in one system, with AI doing more than just transcribing them. I built DialPhone as an AI-native communications platform around that idea. The core is cloud VoIP, while the same platform also handles business SMS, video meetings, online fax, team chat, and contact-center workflows. On top of that, I added AI capabilities such as call transcription and summaries, CRM logging, AI-drafted SMS replies, conversation intelligence, and an AI receptionist that can answer calls, qualify requests, book appointments, and route conversations. The biggest architectural decision was to treat communication data as one connected stream rather than as separate products. A phone call can create CRM context, trigger a follow-up SMS, and become part of a customer-support workflow without someone manually moving information between systems. Along the way, I learned that adding AI to communications is not particularly useful if it only produces transcripts. The more interesting problem is turning conversations into actions: updating records, identifying next steps, helping agents during calls, and automating repetitive work. Step-by-step: 1. I identified the problem of business communication being fragmented across phone, SMS, meetings, fax, and customer-support systems. 2. I built DialPhone as an AI-native communications platform centered on cloud VoIP. 3. I brought business SMS, video meetings, online fax, team chat, and contact-center workflows into the same platform. 4. I added AI features for call transcription and summaries, CRM logging, AI-drafted SMS replies, conversation intelligence, and an AI receptionist. 5. I connected communication data so calls can create CRM context, trigger follow-up SMS messages, and become part of customer-support workflows. 6. I focused the AI capabilities on turning conversations into actions, including updating records, identifying next steps, helping agents during calls, and automating repetitive work.

This Codex-driven workflow takes a software request from problem understanding through implementation, validation, review, and delivery evidence. Instead of asking an AI coding agent to simply “build the feature,” it gives the agent an explicit development lifecycle with defined responsibilities, deterministic validation gates, repair loops, and human checkpoints. The goal is to make AI-assisted development more structured, observable, and recoverable. It can be used for new feature implementation, bug fixing, refactoring, test creation and improvement, code quality and security hardening, and documentation and automation changes. The core principle is simple: Don't give the AI only a coding task. Give it an engineering lifecycle to work on. Step-by-step: 1. Understand the problem. Clarify the request, identify the desired outcome, define the scope, and surface ambiguity before implementation begins. The output is problem understanding and scope. 2. Define constraints. Identify technical, functional, non-functional, compatibility, and out-of-scope constraints. The output is a constraint set. 3. Plan. Analyze implementation options, select an appropriate approach, break the work into tasks, and define acceptance criteria. The output is an implementation plan. 4. Inspect the existing system. Review the relevant codebase, dependencies, current behavior, and affected components before making changes. The output is system context. 5. Implement. Make the smallest appropriate code changes while following the existing project’s conventions and the approved plan. The output is code changes. 6. Run deterministic validation. Run tools that can objectively validate the implementation, including formatting, linting, type checks, builds, unit tests, and other available automated checks. The output is validation results. 7. Review. Evaluate the implementation against the original requirement, the plan, code quality expectations, security considerations, and potential regressions. The output is review findings. 8. Repair and iterate. If validation or review identifies problems, diagnose the issue, make the required correction, and repeat validation. The output is a corrected implementation. 9. Verify. Confirm that the acceptance criteria are satisfied and that the relevant tests and checks provide sufficient evidence for completion. The output is a verification result. 10. Produce delivery evidence and handoff. Summarize what changed, what was tested, what passed, known limitations, and any remaining decisions requiring human attention. The output is delivery evidence and a human handoff. The core loop is: Implement → Validate → Review → Repair → Validate → Verify. Testing and review are treated as part of development rather than activities performed only after coding is “finished.” AI coding agents are increasingly capable of inspecting repositories, writing code, running commands, and responding to failures. The problem is that capability alone does not provide an engineering process. This workflow separates the responsibilities an AI coding agent performs into explicit stages. It applies several principles: - Problem before implementation: Understand what needs to change before writing code. - WHY before HOW: Establish the intent and constraints before choosing an implementation. - Single responsibility per stage: Give each stage a defined purpose and output. - Deterministic validation first: Use tests, linters, type checkers, builds, and other deterministic tools wherever they can establish correctness. - Failure localization: When something fails, identify which stage or assumption needs correction. - Evidence-based completion: Support completion with validation and review evidence rather than an AI declaration that the task is finished. - Human checkpoints: Use automation to accelerate execution without removing human judgment from important decisions. The broader idea is: The AI should participate in the engineering system, not become the engineering system.

Having a team of specialized AI agents creates a new problem: someone still needs to decide which agents should work on a project, what order they should work in, what each one needs from the others, and whether the project is actually finished. Without coordination, the human becomes the project manager, manually moving context and outputs between agents. I created a project-manager agent that acts as the orchestrator for my AI team. I give it an objective, and it determines what work needs to happen, selects the appropriate specialist agents, sequences their work based on dependencies, and presents the execution plan to me before anything starts. Once I approve the plan, it coordinates the agents, manages their handoffs, tracks project state, and maintains enough persistent context for the work to continue across sessions. Step-by-step: 1. I create several specialized agents with clearly defined responsibilities, capabilities, and expected outputs. 2. I create a project-manager agent that knows what each specialist does and is instructed to orchestrate the work rather than perform specialist work itself. 3. I give the project manager a high-level objective. It analyzes the goal, inspects the project context, identifies the required work, and selects the appropriate agents. 4. I have it create an execution plan showing which agents will be used, what each one will do, their dependencies, and the order of execution. 5. I require human approval before execution begins. I can approve the plan, narrow the scope, change the sequence, or redirect the project as needed. 6. Once the plan is approved, I let the project manager delegate each task to the appropriate specialist and pass relevant context and prior outputs between agents through structured handoffs. 7. I track progress and project state as the agents complete their assignments. If an agent uncovers new work, fails review, or changes the project assumptions, the project manager updates the plan and routes the next work accordingly. 8. At the end of the session, I save the current state, completed work, important decisions, and next actions so another session can continue without reconstructing the project from scratch. Instead of personally coordinating every AI agent, I manage the project at a higher level: I define the objective, approve the plan, review important decisions, and evaluate the result. The AI project manager handles the coordination layer, turning a collection of specialized agents into a team that can execute complex, multistep projects coherently.

My AI digest looked perfect and was quietly wrong. What actually fixed it. Every run succeeded. Every dashboard was green. And the content was still wrong. My digest invented "AI Moat Brief", a newsletter that does not exist. It reported scan counts nobody measured. It resurfaced week-old stories as fresh headlines. Here is what broke, and what fixed it. The sorting used to happen in my head: skimmed subject lines, unopened tabs, quiet guilt. The Signal is one email at 08:00: a single Make scenario calling Claude Sonnet through OpenRouter. It reads the last 24 hours of my RSS feeds and newsletters, keeps what touches what I am actively building plus the domains I need to stay current in, and arrives in the language I actually think in. Core items end with what it means for my work. Five to ten minutes, and I know where to go deep today. Structurally it looks like this, minus the content, rendered in English for this post (Image 1). No real edition is shown; section names and sample lines are illustrative. The dangerous failures were never pipeline failures. They were content failures, and the cause is structural: an LLM summarizing newsletters that already summarize primary sources is third hand by construction. Every hop strips attribution and adds confidence, and when data goes missing the model fills the gap the way LLMs do: fluently. Valid HTML, confident tone, green pipeline, wrong content. Image 2 is that whole failure class in one frame. Three rules closed the gaps I caught, all live in production: Step-by-step: 1. Verbatim or nothing. A source name is copied character for character, and a link exists only if that exact URL is in the input. The model copies; it never composes. 2. The model never generates metrics. Every count the report shows is injected by the pipeline after the model returns. 3. Recycled news gets demoted. A recap of recaps gets one line at most, and is dropped when the underlying story falls outside the collection window. Rules 1 and 3 lean on the prompt, and that is why the counters exist. The pipeline writes a hidden HTML comment into every email it sends: items, links, urls, cost, finish status. That line caught what I could not see. In one run, the published-links counter and the leftover-urls counter read 45 and 435: the only sign a new cleanup step was a silent no-op. Another morning the model stopped at 15,999 tokens against a 16,000 cap, one token from an email cut off mid-sentence. On the morning I wrote this they agreed, 21 links and 21 urls, and boring is the goal. Image 3 is that morning's actual comment, with the same two counters from the no-op run. The run itself has a dead man's switch on Healthchecks.io, so a missing 08:00 email reaches me before I notice. Honest limits: the $0.31 per report is a fresh measurement I am still validating, and I have not proven these rules hold as the source set scales. There is more behind every part of this; I would rather share it where it is wanted. Ask and I will put it in the comments: the three rules in full, the exact cost and what drives it, what this replaced in my day, how it compares to what is on the market, or the ugliest of the 15 documented bugs. I am sharing this because I doubt I am the only one building fragile things behind the scenes, and monitoring text is harder than monitoring uptime. What content-level checks do you run on LLM output, the kind pipeline monitoring cannot see? Real thresholds and embarrassing failures especially welcome.

AI agents can generate impressive work quickly, but the agent that created something is not necessarily the best judge of whether it is correct, complete, secure, or ready to move forward. Without an independent review step, mistakes can compound as later stages build on work that was never properly validated. I created a system of specialized AI review agents that act as quality gates between stages of work. Instead of letting the agent that performed the work decide whether it is finished, a separate reviewer evaluates the output against explicit criteria and makes a gate decision: PASS or NEEDS REVISION. Different reviewers focus on different dimensions. In my software development workflow, I use reviewers for implementation fidelity, code quality, security, performance, and specification compliance. A feature does not advance until the required reviewers have passed it. Step-by-step: 1. Define what “good” means before the work starts. Give reviewers an explicit source of truth, such as a specification, plan, acceptance criteria, coding standards, security rules, or quality rubric. 2. Separate execution from evaluation. The agent that performs the work should not be the only agent deciding whether that work is acceptable. 3. Create specialized reviewers for important quality dimensions. For software, this might include implementation, code quality, security, performance, and specification reviewers. The same pattern can be used for research, writing, factuality, compliance, financial analysis, or brand review. 4. Run the appropriate reviewers when a stage is complete. Each reviewer independently inspects the work from its assigned perspective and actively looks for reasons it should not advance. 5. Require an explicit gate decision. A reviewer must return either PASS or NEEDS REVISION, along with concrete findings and recommended fixes. In my workflow, reviewers can block progression for issues such as missing tests, even when the underlying implementation appears correct. 6. Route failed work back to the appropriate agent. The worker fixes the identified problems and submits the work for review again. 7. Advance only after the required gates pass. Later stages should not build on work that still has unresolved review findings. 8. Keep humans at consequential decision points. AI reviewers can determine whether work satisfies their assigned criteria, but important actions such as merging, deploying, publishing, or otherwise committing the result can remain human decisions. Instead of treating AI-generated work as complete simply because an agent produced it, I create a controlled loop: Build → Review → Fix → Re-review → Pass → Advance The result is a more reliable workflow where specialized agents perform the work, independent agents challenge it, and errors are caught before they propagate into later stages.

I built AllegraMente, a complete, free iOS cognitive training app, solo with Claude Code. It includes 11 exercise areas, 83 articles in five languages, and no tracking. My workflow starts with a new feature or content idea, a new exercise type, or an in-depth article about how memory works. Planning and execution stay separate, and the instruction document serves as the contract. Because everything is specified upfront, Claude Code almost never goes off track. Step-by-step: 1. I plan the feature with Claude in chat, define the data models and UX, and fact-check every scientific claim against primary sources before writing anything. 2. I package the result into one self-contained Markdown instruction document containing file paths, specifications, localization keys for all five languages (IT/EN/ES/FR/DE), edge cases, and acceptance criteria. 3. I create a Git restore point. 4. I give the document to Claude Code, which implements the feature end to end using SwiftUI, SwiftData, and MVVM. 5. A small script sends me a Telegram notification when the task is complete. I review the diff and test the app on a device. 6. I ship by creating an archive with `xcodebuild` and uploading it to the App Store with Transporter. This process is how I shipped 11 exercise areas, 83 articles in five languages, and a complete editorial system as a solo developer.

After about 18 months of building software with AI, I realized that reliability wasn't primarily a model problem. Bigger context windows and more clever prompts didn't fix it. What did help was treating the AI like a developer joining an existing team instead of like a chatbot. Real developers don't work from memory. They inspect production, read the documentation, check the tickets, and use proven tools. I built an environment that lets the AI do the same. The workflow is tool-agnostic, so it can be rebuilt with whatever AI client and stack you already use. Step-by-step: 1. I gave the AI a persistent task and history store that it can read from and write to. This is the core of the workflow. Mine lives behind an MCP tool, but any queryable store can work. Every architectural decision, blocker, and progress note gets written there instead of being left in the chat. 2. I open every session with a stand-up. Before writing a single line of code, the AI pulls what was in progress, what's blocked, what changed since the last session, and which architectural decisions still hold. About 30 seconds later, we're both looking at the same project. Then we build. 3. I exposed real operations as MCP tools instead of relying on "write code" prompts. I wrapped specific, tested actions—such as creating a page, defining a data model, wiring an integration, and running a migration—as tools. The AI composes these known-good building blocks into larger solutions instead of regenerating infrastructure every session. I call this wave coding, and it's the biggest reason the output stays consistent. 4. I made verification a rule: before touching anything, the AI reads the live database, API state, logs, and files. It checks ground truth first instead of making assumptions. 5. I made the chat disposable and the log canonical. If it isn't logged, it didn't happen. The task store is the single source of truth, not the conversation. The payoff is that I can stop halfway through a feature, close my laptop, and come back days later. The AI reconstructs the project from its own history, so I don't spend 20 minutes re-explaining it. Full disclosure: I built this into my own platform, WebsitePublisher.ai, which currently has 43 MCP tools and 105 integration building blocks. It's delivered as an add-on that plugs into the AI client I already use over MCP, so there's no new app to learn. Nothing here is locked to that platform, though: the workflow itself works with any MCP client and any store the AI can query. I'm curious whether anyone else is running their AI this way or solving the amnesia problem from a different angle.

A finished video can be wrong in ways you cannot see—not because of the edit, but because of the delivery file itself. That is what gets work sent back, and it is rarely the craft. I built a delivery check for this. I give it a finished file, tell it where the file is going and what kind of piece it is, and it measures the things that cause rejections: sample rate, mono audio, integrated loudness against the destination target, true peak, dynamic range against a band rather than a single number, A/V drift, whether the shots cut together, and whether the file can stream before it has finished downloading. Most of those checks are for sound, because most of what gets sent back is sound. Picture problems are visible on a screen. A file that is 2 dB too quiet or peaks at -0.7 dBFS can look perfect and still come back. For everything it can fix, the checker gives me the exact `ffmpeg` command, with the numbers already calculated for that file. It does not merely describe the fix, and it does not hand me a corrected file. That was the decision that mattered. A fixer is a black box. You never learn that you had a problem, so you make it again the following week. Then, when your editing tool adds an “optimize on export” button, you have nothing. An inspector that explains the problem in one sentence and gives you the command teaches you the standard once and remains useful when the tools change. Three things determined whether I would actually use it, and none of them are checks. “Two severities, never one.” Something either bounces, or it needs your eyes. A tool that only says “bad” gets ignored on the third run, because half of what it flags is a decision you made on purpose. Sample rate, mono, and true peak bounce without argument, and they get a command. Loudness and dynamic range need to be reviewed first. My dynamic-range check says in plain words that the result is often deliberate and should be fixed at the source rather than in the master. “A check that refuses to give a verdict.” On vertical video, the platform interface covers the bottom 26 percent and the top 12 percent. I flag high-contrast elements in those zones but deliberately do not fail them, because the detector cannot tell a caption from a bright patch of sky. The report explains that limitation. A check admitting what it cannot know is what makes the checks that do commit worth trusting. “The thresholds are mine; the code only applies them.” They live in a table: destination crossed with content. Social wants -14 LUFS, a festival master wants -18 with a wider tolerance because festivals provide a band rather than a number, and broadcast wants -23. A scripted short is allowed to be denser than a screencast. That table is the whole product. For example, to catch shots that do not cut together, I measure the luminance range across the piece. The first version used average brightness per frame and kept flagging legitimate night photography. A fade to black has nothing bright in it; a night scene does—a streetlight, a moon, or a face. Changing the discriminant to the brightest pixel in the frame instead of the average eliminated the false positives. That took ten minutes of thinking, and no amount of better code would have found it. Step-by-step: 1. I wrote down what had actually gotten my work sent back over twenty years before writing any code. I captured the scars rather than making a spec sheet; that list became the product. 2. I gave every check a severity: it bounces, or look before you send. Anything I could not confidently put in one bucket became informational, with no verdict at all. 3. I made every check return four things: the measured value, pass or fail, why it matters in one plain sentence, and, where possible, the command that fixes it. 4. I put the thresholds in a profile table instead of hardcoding one standard, because -14 LUFS is right for social and wrong for a festival. 5. I added a parameter for whether the file is my own master or a copy pulled from a platform. On a downloaded copy, half the container checks measure someone else’s transcode rather than my work, so they are skipped and the report explains why. 6. When two fixes would collide, I output only one command. If loudness already needs a gain change, the limiter goes inside that command instead of being offered separately; otherwise, I would run two instructions that fight each other. 7. I tuned the checker against real files until the false positives stopped. Ignoring the top and bottom five percent of frames eliminated the ones caused by a single stray frame. 8. I made the output a report. I read it, decide, and run the command myself. The functions are an afternoon of work, and anyone can copy them. What is not written down is the list of what to check, at what threshold, and why.

I use AI agents across roughly half of my startup, including search, security testing, development, and affiliate integrations. - Smart Shopping Deals search: It never relies on a single AI provider. I use a chain of backup providers for LLM-based search and embeddings with pgvector. If one provider fails or times out, the system silently retries with the next. A slow provider never hangs the request, and users never see the failure. In the worst case, the system falls back to plain search. - Security testing: I run the autonomous AI security agent Strix against my live preview after every deploy. It actively attacks the app the way a hacker would. - Development workflow: My real “team” is a four-layer testing rule enforced by the AI itself. Every feature has to ship with backend tests, component tests, full browser end-to-end tests, and test-data setup, all in the same commit. Claude Code doesn’t consider a feature “done” until all four exist—not just the code. - Affiliate parsing: One parser handles every affiliate network. I paste in the raw ad code from any network, and the same parser automatically extracts the banner link, image, and destination URL. No network-specific code is needed.

I created an agent collaboration system called duo-agents that pairs Claude and Codex to work together on coding tasks... Claude acts as the implementer (coder), then Codex acts as the reviewer (checks and makes edits). They alternate in rounds, communicating through a shared file. The key difference: both agents actually edit files — the reviewer doesn't just leave comments, they make the fixes themselves. Describe your task and watch them iterate until the code is solid. Step-by-step: 1. I created a shared file that both coding agents could use to pass context and decisions back and forth. 2. I assigned Claude the implementer role and had it build the requested change directly in the codebase. 3. I assigned Codex the reviewer role and had it inspect the implementation for problems. 4. Instead of leaving comments, Codex edited the files and made the fixes itself. 5. I alternated the two agents in rounds until the shared task was complete and the code was solid.

I developed a platform that helps users learn dimensional modeling and create entity-relationship (ER) diagrams through a drag-and-drop interface tailored to specific use cases. I noticed that many people struggle with visualization and face a significant gap between understanding database concepts and creating tables. This can make it difficult to work independently or determine which type of table to use and when, including whether to create a dimension or fact table. Step-by-step: 1. I identified the difficulty many people have visualizing dimensional modeling concepts. 2. I created a platform focused on learning dimensional modeling and ER diagrams. 3. I built a drag-and-drop interface tailored to specific use cases. 4. I designed it to help users understand how to create tables and determine whether to use a dimension table or a fact table.