ClaudeHack
0.000

Audit for instruction collision between CLAUDE.md, skills and your prompt

Claude Code / CLAUDE.md·15 hours ago·verified 15 hours ago·easy·~20 min setup·verified

"Never add comments" in CLAUDE.md plus "make this readable" in your prompt is a contradiction Claude has to resolve before it writes a line. Find those pairs and pick one.

The problem

Instructions pile up in three places — the system prompt, CLAUDE.md and skills, and the prompt you just typed. When they disagree, reasoning tokens go on reconciling your own rules instead of on the task.

The hack

Audit your instruction layers for direct contradictions. Do this whenever you switch to a newer model, rather than only swapping the model and assuming the rest still applies.

Why it works

Rules written as guardrails for an older model become hard limits on a newer one. A contradiction does not just waste tokens — it makes behaviour unpredictable, because which instruction wins depends on wording and position rather than on your intent.

Setup

1. Run `/claude-api prompt-audit` to flag instructions written for older models across your prompts, skills and tool descriptions, and get the fixes as a diff. 2. Read `CLAUDE.md` alongside your most-used skills and list any rule stated twice. 3. For each duplicate, keep it in one place only — tool guidance belongs in the tool description or in `CLAUDE.md`, not both. 4. Replace absolute rules with judgement where you can: "write code that reads like the surrounding code" instead of "never write comments". 5. Re-run a task that was behaving oddly and see whether it settles.

Prompt

Read my CLAUDE.md and every skill in .claude/skills/. List any pair of instructions that contradict each other, and any instruction that appears in more than one place. Do not change anything yet.

Additional details

The classic three-way collision looks like this: the system prompt says not to add comments, `CLAUDE.md` says to document everything, and your prompt asks for readable code. All three are reasonable on their own. The related fix is structural rather than editorial. A 1,400-line `CLAUDE.md` where the testing section matters in one session out of twenty still loads in full every time. Move the rarely needed sections into skills that load when they match, and let `CLAUDE.md` point at them — it works better as a router to a tree of files than as the file that contains everything.

Source

youtube·View original →·by Eric Tech

Does this still work?

0 said it works · 0 said it's broken

Related

Discussion

No comments yet. Be the first to say whether this worked for you.