Find the bloat in yourCLAUDE.md
Your CLAUDE.md is re-sent on every turn, 2 prompts or 200, competing for the same context budget that drives output quality. ccmd marks up the file you already wrote line by line, tags the rule each finding breaks, and puts a token and dollar cost on every one. Same scoring for AGENTS.md, .cursorrules, and .grokrules.
Anthropic announced a separate, capped Agent SDK credit pool for claude -p and headless runs, then paused it on June 15 after pushback. For now agent usage still draws from your normal Pro/Max limits, and Anthropic says a reworked version is coming with notice. The cost that never paused: your CLAUDE.md is re-sent on every turn and competes for the context budget that drives output quality.
Marked-up diff, per-line callouts, and a token cost in ~50 seconds. No signup, nothing to install. No file in hand? Hit See example report to read a real one first.
Your file goes to our Opus 4.7 analyzer and is logged with the rewrite so we can keep the rubric honest; reply "remove" on any email and we delete your record.
Claude Code shipped a memory primitive (CLAUDE.md) and gave it almost no guardrails. Power users figured out the hard way that the file is sent on every turn, that it competes for the same context budget that drives output quality, and that a sloppy 8k-token file silently taxes every prompt for the rest of the session.
every single API call to Claude sends the whole context, including prompts, meaning that all this extra text in CLAUDE.md is sent over and over.— caymanjim, HN 47581701
Anthropic then announced it would split programmatic Claude Code billing into a separate, capped Agent SDK credit pool, paused the change on 2026-06-15 after pushback, and said a reworked version is coming with notice. Billed together or apart, every token your config re-sends is cost you pay again on every turn.
Existing tools split into generators (paste a stack, get a starter) and cost meters (tell you what you spent). Neither audits the file you already wrote. ccmd does.
Same rubric across four agent ecosystems. Polyglot from day one so you can audit every config in your repo without learning four tools.
- claude.mdAnthropic Claude Code memory file.
- agents.mdOpenAI Codex / GPT-5.5 agent contract.
- .cursorrulesCursor IDE rules; .mdc multi-file variant supported.
- .grokrulesxAI Grok Build CLI (May 2026). Native parsing.
Four rules, distilled from Karpathy by Forrest Chang. The community measured a 41% → 11% mistake-rate drop.
ccmd's default rubric scores your file against four: Think-Before-Coding, Simplicity-First, Surgical-Changes, Goal-Driven-Execution. Every callout in the analyzer above is tagged with the rule it triggers (or null when the issue is plain bloat). The hero panel renders the side-by-side rewrite live; read the long version.
ccmd is free. The analyzer is fully working above. Sign up to get more.
The in-browser analyzer is the product, free forever, no signup required to use it. Sign up to be onboarded into the features I'm shipping over the coming weeks:
- ·GitHub webhook on push: your config gets re-scored every commit
- ·Weekly drift email with a real diff and a token-savings number
- ·Session-trace analysis: which rules actually fire in your Claude Code sessions
- ·PR diff comments on every CLAUDE.md change
- ·Per-engineer Agent SDK cost rollups for teams
Why not just ask Claude to optimize my CLAUDE.md?
Claude pasting its own file back at itself is a 200-token tip. ccmd measures: which lines fire, which never do, which contradict the system prompt, and what every line costs at Opus 4.7 rates. A model cannot count its own tokens reliably and cannot see your billing.
Does it work with Codex, Cursor, Grok Build?
Yes. Polyglot from day one. AGENTS.md (Codex), .cursorrules (Cursor IDE), and .grokrules (xAI Grok Build, shipped May 2026) all use the same scoring rubric. One tool for every config file in your repo.
How is this different from Anthropic's claude-code-setup plugin?
claude-code-setup recommends structure for a brand-new project. ccmd grades the file you already have, against a rubric tied to specific past failure modes, with a token-cost number on every finding.
Is it really free?
Yes. The analyzer above is free to use; your file is sent to our Opus 4.7 endpoint, scored, and the rewrite + callouts stream back. We log the upload and the rewrite alongside an IP hash so we can iterate on the rubric and answer support emails; reply "remove" to any email and we delete your record. I'm shipping a paid tier with continuous monitoring + GitHub webhook + drift email + team rollups, and you can sign up to be onboarded as those land. The free analyzer itself stays free forever.