AI usage audit + meta-review prompt

A developer conducted an audit of their own AI usage by instructing Claude Code to analyze configuration files, prompt history, and transcripts across their system. The meta-review aimed to evaluate the developer's prompt engineering and context management against current best practices, requiring the AI to fetch fresh documentation and web sources rather than relying on stale training data. The audit produced a prioritized build list for improving skill/agent ROI, reducing duplicate prompts, and fixing broken configurations.

Copy everything below into a Claude Code session started from your home directory or wherever your config lives . Works best on a strong model with web access. Audit how I actually use you, then meta-review my prompt & context engineering against the current state of the art. Use your own knowledge of how Claude Code, skills, agents, hooks, memory, and model routing are meant to work as the standard I'm measured against. Everything you claim must trace to something you read this session — no advice that would apply to any random user. Do not write or edit any files; end with a prioritized build list and let me pick what you implement. - Find my config: ~/.claude/ and $CLAUDE CONFIG DIR if set; I may have more than one — e.g. a separate work volume. Ask me if unsure, and if a second config dir is unreachable, say so and audit what you can . - In each config dir, read: CLAUDE.md , settings.json , every skills/ /SKILL.md frontmatter + skim of the body , every agents/ .md , every hook script referenced from settings, and the memory directory if present. - Parse each history.jsonl IN FULL — it is every prompt I've ever typed. Extract with jq : prompt volume per month, per project, slash-command frequency, and exact-duplicate prompts normalize case/punctuation, count anything typed 2+ times . - Transcripts under projects/ are usually too large to read raw. First check the retention window oldest transcript mtime and STATE it — any "never invoked" claim must be scoped to that window. Then:- Grep the FULL transcript corpus mechanically for: skill launches "skill":"..." and Launching skill: , <command-name tags, subagent types "subagent type" , tool-use rejections "doesn't want to proceed", "user rejected" , and model ids "model":"claude-..." . Never sample for these — counting is cheap and sampling lies. - For qualitative signals, extract the user messages from the ~20 most recent sessions plus the ~5 largest ones into a scratch file jq 'select .type=="user" ...' , and delegate reading/synthesis of that file to subagents. Run all shell extraction yourself from the main session. - Grep the FULL transcript corpus mechanically for: skill launches - Verify any docs-dependent claim hook input format, frontmatter behavior, settings semantics by fetching the current docs at code.claude.com/docs — do not assert from memory. - What do I actually use you for — by project, task type, and time? Has that shifted, and does my skill/agent/config investment match the CURRENT distribution or a past one? - What do I retype that should be config? Exact-duplicate prompts, per-session re-explained context build commands, layouts, credentials, known issues , and repeated corrections of the same mistake. Quote each verbatim with its count. Flag any secrets I've pasted into prompts. - Skill/agent ROI: for each one, invocation count history slash + transcript launches, window-scoped vs maintenance cost. Which are dead weight, which have broken triggers missing/weak frontmatter descriptions , which are redundant pairs? Judge by measured use, not file quality. - Where am I fighting the tool: permission denials and what was being denied , interrupted/abandoned sessions, context-exhaustion chains, restarts that re-paste the same ask, hooks that silently never fire test them against the documented contract . - Model routing: what do I run by default vs what the tasks need? Count model usage from transcripts, note manual /model / /effort switching, and say what cheaper routing would need skills, CLAUDE.md lines, better briefs to keep ~90% of the quality. My prompt & context engineering, graded against the current state of the art — and "current" means you must do fresh web research first this field moves monthly; your training data is stale by definition : - Pull 5–10 recent, reputable sources on context engineering, agentic coding workflows, spec-driven development, agent evals, and multi-agent orchestration vendor engineering blogs, practitioner state-of-the-field posts — not SEO listicles . - For each major trend: am I ahead, current, or behind — with evidence from my own artifacts and transcripts quote or path per verdict . Call out anything I converged on independently before it was a named trend, and anything I hand-rolled that the platform now ships natively. - What to learn, in order of leverage: 3–6 items, each with the specific resource to read AND the first concrete exercise to do in my own setup not "learn about X" . - One-sentence summary of my strongest half and my growth edge. One report, two parts, 2,500 words max total, with a Sources list of every URL used: Part 1: doing well / doing wrong ranked by impact / concrete optimizations each names the exact file to create or edit and what goes in it . Part 2: the meta-review as above. End with a prioritized build list of the artifacts you'd create. Then STOP — implement nothing until I pick. When I do approve items, verify each change after making it run the hook, re-grep the state, launch a throwaway session if needed rather than assuming it landed.