This is the exact prompt I use to turn a tool I don't know into a skill my AI does. Point it at a web app OR an API. It researches the whole thing end-to-end (READ-ONLY), then builds a clean, reusable Claude skill so the AI can operate it — or answer anything about it. Run it on ULTRA effort and let it fan out DYNAMIC WORKFLOWS (parallel subagents) — that's how the research gets genuinely exhaustive instead of a skim. Fill the two brackets, paste it into Claude Code, and let it cook 👇 ——— THE PROMPT ——— Goal: build me the most comprehensive Claude SKILL for operating <TOOL / PLATFORM / API> end to end. TARGET: <the web-app URL —OR— the API name + its docs URL> ACCESS: <"my logged-in browser via /chrome" —OR— "API key in .env"> Use ULTRA effort. Do EXHAUSTIVE deep research FIRST — fan out DYNAMIC WORKFLOWS (parallel subagents, one per module / endpoint / doc section), then synthesize — then build the skill with /skill-creator. RESEARCH — understand it completely, end to end: • Web app: open it in my browser and walk EVERY section, tab, module, and setting. Map the whole thing — navigation, every feature, how workflows/automations are built, the AI features, integrations — and note WHERE each capability lives in the UI, so the skill can later click straight to it. • API: read ALL the official docs. Enumerate every endpoint/resource, the auth model, rate limits, pagination, webhooks/events, SDKs, error shapes, and the common end-to-end workflows. Capture the gotchas. • Either way: read their documentation in full and KEEP THE LINKS — the skill should point to the exact doc for anything it can't hold in memory. READ-ONLY (critical): do NOT modify, create, delete, send, or change ANYTHING in my account or data. Don't trigger automations, don't edit settings, don't remove anyone. Only observe and document. Treat it as production. BUILD THE SKILL (use /skill-creator + best practices): • Comprehensive but CLEAN and modular — a short SKILL.md that ROUTES to a references/ folder. A hundred files is fine; organize them so the AI can find any answer instantly (load only what a task needs).

I’m a big fan of Claude Code and the work behind it. Recently, though, I’ve been using Codex and Nano-Banana more often, especially for content creation and my day-to-day work. I’ve found these models to be more capable overall. Codex, for example, reliably follows instructions and tends to be more analytical and careful in its responses. I’m not discouraging the use of tools like Claude Code, but I do recommend exploring other options as well. In my experience, Claude Code feels more creative, while Codex is stronger for tasks that require analytical precision and consistency. I’ve built skills around these tools and now rely on Codex for more structured or detail-sensitive work. I expect these models will continue to improve, and Codex also offers better rate limits, which makes it especially useful for my workflow. I’m also a big fan of ChatGPT Pro and have been using it extensively. A lot of my workflow is prompt into chatgpt pro extended thinking -> codex implementing. These are just some quick notes I wanted to share about the different models I’ve been using. This was recorded with my locally developed Whisper system, using Codex — a nice full-circle moment. I’m hoping to schedule a video call next week with others to discuss these topics further. I am also thinking of making a few classroom modules and have some ideas (creating and using skills / connecting up all your data), I have a friend obsessed with "Jarvis" (2 of them actually...). But although "Jarvis" is kind of larp, the idea of having one center station makes sense. This may be a topic we talk about on a call or a classroom module :P. Have a great start to your week :)