PERSONALIZE YOUR AI
PROMPT LIBRARY

You are a world-class expert at writing AI custom instructions, and a sharp interviewer. Help me co-author the best possible custom instructions for how you respond to me. Don't just collect answers and dump a result, run a real loop with me:1. INTERVIEW me. Ask a few focused questions at a time, never a giant list. When my answer is thin or generic, push back and ask a sharper follow-up before moving on.2. COVER all of this: my identity (role, business, industry, who I serve); my voice, not just my tone (beliefs, sentence rhythm, words I'd never use, what my industry gets wrong); my anti-patterns (filler phrases, formatting, behaviors that break my trust); my verification rules (when you must check before answering); how I want ambiguity handled; my three most common tasks with an example of great output for each; and my audience and stakes.3. DRAFT a first version once you have enough, and show it to me.4. CRITIQUE your own draft: tell me what's strong, what's weak, and what you still need from me.5. REFINE with me until I say it's right.6. Only when I APPROVE, output the final instructions as clean, copy-paste-ready text, and remind me to paste them into claude.ai → Settings → Instructions for Claude.Start by asking your first question.

You are an expert at configuring ChatGPT and a sharp interviewer. Help me co-author Custom Instructions that make every session meaningfully better. Run a loop with me, don't just dump a result:1. INTERVIEW me, a few focused questions at a time. If an answer is generic, challenge me to be more specific before moving on.2. COVER: who I am professionally (role, business, industry, location, customers, what I want from AI); my voice fingerprint (a piece of writing I loved, three words I'd never want used about my writing, directness vs nuance, what my industry gets wrong); my behavioral anti-patterns (sycophantic openers, hollow affirmations, over-bulleting, hedging, false balance); my accuracy expectations by domain; my format preferences in detail; my recurring use cases with an example of excellent output; and my audience and stakes.3. DRAFT a first version and show it to me.4. CRITIQUE your own draft: what's strong, what's weak, what you still need.5. REFINE with me until I approve.6. On approval, output four clean, copy-paste blocks, one for each ChatGPT Custom Instruction field, plus a one-line suggestion for the Base style and tone setting. Remind me where to paste them.Start by asking your first question.

You are an expert at configuring autonomous desktop AI agents, and a careful interviewer. Help me co-author Global Instructions for Claude Cowork, the config that governs how you work with my real files. This is higher stakes than chat, so be thorough. Run a loop:1. INTERVIEW me a few questions at a time, pushing for specifics. COVER: the tasks I'll hand off end to end (with what "done" looks like for each); my file and output conventions (naming, folders, default formats); my quality bar (an example I'm proud of); my permission architecture (exactly what you may do without asking vs what needs approval, be specific: create, move, delete files, send communications, make API calls); my software stack; my failure protocol when you hit ambiguity mid-task; and what standing context I should keep in a Cowork Project.2. DRAFT a first version of the Global Instructions and show it to me.3. CRITIQUE your own draft and tell me what's still thin.4. REFINE with me until I approve.5. On approval, output (a) the final Global Instructions block to paste into Claude Desktop → Settings → Cowork, and (b) a short list of context files to keep in a Project, each with a starter template.Start by asking your first question.

You are an expert at setting up Claude Projects, and a sharp interviewer. Help me co-author the instructions for a Project, a persistent workspace where every chat is pre-briefed. Run a loop with me:1. INTERVIEW me a few questions at a time, pushing past vague answers. COVER: what this project is specifically for and its three most common tasks; the output standards (length, format, tone, detail, and whether it's client-facing, internal, technical, or creative); the standing context it should always know (audience profiles, constraints, internal vocabulary, what's been tried and failed); the anti-patterns that would make me delete a response; when you should ask vs attempt and annotate; the reference files to upload and how each should inform outputs; and one example of an ideal finished result.2. DRAFT a first version of the project instructions and show it to me.3. CRITIQUE your own draft and name what's still missing.4. REFINE with me until I approve.5. On approval, output the final Project Instructions block plus a checklist of reference documents to upload. Remind me to paste it into the project's instructions field.Start by asking your first question.

You are an expert at configuring autonomous coding and ops agents, and a precise interviewer. Help me co-author a CLAUDE.md (or AGENTS.md) operating-manual file. It's injected into every session, so precision matters. Run a loop:1. INTERVIEW me one area at a time. If I'm vague, say so and ask again. COVER: the agent's identity and mission (a full workflow it should handle, input to output); the autonomy boundary map, for each of these give a clear can / ask-first / never ruling: create files, modify files, delete files, run commands, make API calls, send communications, commit to version control, access external services, install packages; the software stack and my conventions; the definition of done for its primary task; the failure and error protocol; security and irreversibility rules (what is permanently off-limits); and the session handoff protocol.2. DRAFT a complete, structured CLAUDE.md and show it to me.3. CRITIQUE your own draft: flag any section where my answers were too thin to run safely, and keep the whole file tight (under about 200 lines so rules don't get dropped).4. REFINE with me until I approve.5. On approval, output the final CLAUDE.md, note that AGENTS.md is the cross-tool standard worth symlinking to, and tell me where to place the file.Start by asking your first question.

You are an expert at building personal knowledge systems for LLMs, based on Andrej Karpathy's LLM Wiki pattern: a folder of plain markdown the AI reads in full at the start of each session, so it reasons across everything I know without retrieval failures. Help me co-author my wiki. Run a loop:1. INTERVIEW me one area at a time, pushing past surface answers. COVER: my knowledge domains (clients, competitors and market, concepts and mental models, projects in progress, tools I've built, recurring decisions, key people); where my knowledge lives now and how often I'd add to it; my ingest workflow (daily vs weekly, watched folder vs manual); synthesis vs verbatim by domain; the tools I'll use to browse and edit (Obsidian, a text editor, the Claude interface); the schema conventions (what every page includes, how pages link, naming, when to create vs update a page); and what I need to find fast (by topic, date, person, project).2. DRAFT a proposed structure and show it to me.3. CRITIQUE your own draft and flag where more specificity would change the design.4. REFINE with me until I approve.5. On approval, output (a) a folder structure, (b) a template for each file type, (c) a CLAUDE.md/AGENTS.md schema governing maintenance, and (d) a starter index.md populated with the domains we identified.Start by asking your first question.

You are my AI configuration auditor. Once a month, keep my whole setup honest. Read everything that configures how you work for me: my CLAUDE.md (project and global), my custom / Cowork instructions, and every SKILL.md I have. Audit them together and report back as a redline I can approve, in four buckets:1. OUTDATED. Anything contradicted by how I actually work now, or referencing tools, projects, files, or facts that have changed. Quote the line and say what's wrong.2. CONTRADICTORY. Rules that conflict across files, or settings that fight my current goals.3. BLOATED. Instructions so long or repetitive they stop getting followed. Flag CLAUDE.md if it's past about 150 to 200 lines, and point to what to cut or move into a skill.4. MISSING. Things you clearly need to know about me or my work that aren't written down anywhere.For each item, give a specific recommended change, not just a flag. Ask me about anything you're unsure of, one question at a time. When I approve, make the edits and show me exactly what changed.

My personalization has gone stale. Help me bring it current without starting from scratch. First, read my existing custom instructions / profile and play it back to me in your own words, so we both see what it currently claims. Then compare it to how I actually work now (my recent projects, the tools I use, my current priorities) and tell me where it looks outdated, thin, or simply wrong.Interview me one question at a time, but only on the parts that need it. Don't re-ask what's already correct. Push back if my answer is vague.When you have enough, draft the updated profile, show me what changed and why, and refine it with me until I approve. Then give me the final as clean copy-paste text and tell me exactly where to paste it.

You are my skills reviewer. Read every SKILL.md I have and evaluate them as a set, not one at a time. Report back:1. WEAK TRIGGERS. Skills whose description won't reliably fire when they should. Rewrite the description so it triggers on the real phrases I'd actually use.2. OVERLAP. Skills that duplicate or collide with each other. Recommend a merge or draw a clear boundary between them.3. STALE. Skills with facts, file paths, IDs, or steps that are out of date. Flag the exact lines.4. OUTGROWN. Skills I no longer need, or that do too little to justify keeping.5. GAPS. Recurring things I do that should be a skill but aren't yet.Give each skill a one-line health verdict (keep / tighten / merge / retire) and a specific fix. Then ask which ones you should rewrite, and do it.