The creator of Claude Code says to throw your .md files away every six months. I ran a three-prompt audit instead: 11,000 tokens of always-loaded instructions dropped to 5,800, ten of 91 skills got archived, and nothing I cared about broke. These are the exact prompts.
→My setup was a 134-line CLAUDE.md plus 91 custom skills, and every session started by loading 11,000 tokens of it before I typed a word.
→Anthropic cut 80% of Claude Code's own system prompt when the Claude 5 models shipped. Over-prescriptive instructions now hurt output, and the ALL CAPS rules I wrote to babysit older models were dragging the new ones down.
→The fix was mostly moving knowledge to load on trigger instead of always. Ten things got archived, everything else stayed, and one command restores any of it.
“The bloat came from knowledge loading at the wrong time.”
🔒Drop your email, get the audit prompt pack
The three prompts that ran the whole audit: pull Anthropic's current model docs, grade every rule in your CLAUDE.md and skills against them, then run targeted edits with a full backup. Plus what I learn shipping agent systems every week.
Join 3,000+ builders already getting this. No spam.
THE AUDIT PROMPT PACK — 4 STEPS
01
Pull the current model docs
Paste this into Claude Code, or any Claude with web access. It grounds the audit in what Fable, Opus, and Sonnet 5 actually need now instead of folk wisdom from 2024. All three official per-model prompting guides are linked in the sources below.
Research the current official Anthropic prompting guidance for the
models I actually run (Claude Fable 5, Opus 5, Sonnet 5). Pull the
model pages and prompting docs, plus anything on how CLAUDE.md files
and skills should change for this generation. Summarize what these
models now handle on their own that older models needed explicit
instructions for.
02
Grade your setup against them
The heart of the audit. Four verdicts per instruction: keep, move, archive, soften. The 45-day usage scan kills the guesswork about which skills are actually dead. And the no-overhaul line matters: targeted cuts beat rewrites.
Audit my CLAUDE.md and every skill in ~/.claude/skills against that
guidance. Classify each instruction: KEEP (still earns its place),
MOVE (real knowledge that should load on trigger, not in every
session), ARCHIVE (dead, duplicate, or superseded), SOFTEN (ALL CAPS
or over-prescriptive wording the new models handle without
babysitting). Also scan my last 45 days of session logs and report
which skills actually fired. I do not want a huge overhaul: propose
targeted edits only, and never touch fragile exact commands like
ffmpeg chains or deploy sequences.
03
Run the edits, backed up
Nothing gets deleted without a restore path. My whole run stayed reversible with one tarball, and that backup earned its keep the same afternoon when an automated edit broke 21 files.
Back up my CLAUDE.md and skills to a dated archive folder first,
with a one-command restore. Then apply only the approved edits.
Verify every file still parses, every trigger phrase survived, and
nothing else changed. Show me a full changelog when you are done.
04
What to expect
My numbers from one afternoon: 11,000 always-loaded tokens down to 5,800, ten of 91 skills archived (one duplicate, six dead pointers, three superseded builds), 52 skill descriptions trimmed to clean routing rules, zero regressions since.
Re-run the pack every time a new model family ships. The docs change, and your files quietly go stale.
