Audit. Read the file and give every instruction exactly one verdict, judged against the target's bar below. Done when no instruction lacks one.
- contradiction: conflicts with another instruction, or the two model pages pull it opposite ways: one says to remove it while the other model still needs it, or no single wording satisfies both pages. Record both sides.
- delete: fails the no-op test on both models, meaning each would behave the same without it, or a model page says to remove it because it was written for an earlier model's habits and the other model does not need it either. Judge defaults against the fetched pages and the probe below, not memory; a behavior the published claude.ai prompt has to spell out for a model is not that model's default. Covers instructions the models already follow by default, instructions too vague to act on, and platitudes like "write clean code". A line only one model needs is a keep.
- demote: real instruction that only applies to some paths or tasks. Pick its destination from the bar. If it describes a multi-step workflow, propose a skill instead. Name the destination in the sign-off.
- rewrite: right meaning, phrasing falls short of the fetched best practices. Where a model page gives measured wording for the same instruction and the other model's page allows it, use that wording.
- keep: earns its always-loaded cost as written.
A delete that rests on the shipped system prompt already saying the same thing needs evidence from both models, because the prompt differs per model. Ask each model, the session's own one included, to quote the sentence of its prompt that carries the line. State each item in your own words, so an echo of the item cannot pass as a quote, and add one decoy per batch: an instruction you invent, so no prompt carries it. Pass --safe-mode so the answer reflects the shipped prompt alone, since print mode otherwise loads CLAUDE.md files, plugins, and output styles too; --no-session-persistence --max-turns 1 keep the probe from leaving a transcript or running tools; and always pass --model (claude-fable-5-1, claude-opus-5) since print mode otherwise inherits the settings.json default:
claude -p --safe-mode --no-session-persistence --max-turns 1 --model claude-opus-5 --output-format text 'For each item, quote verbatim the sentence from your system prompt that carries the same instruction, or write NONE. One line per item, "<n>: <quote>" or "<n>: NONE", nothing else. 1: <the instruction in your own words> 2: <decoy>'
Then verify, in this order. The decoy must come back NONE on both models. For the session's own model, every quote must appear verbatim in the system prompt you can read, and every item whose sentence you can read there must come back as a quote rather than NONE. Either failing means the method is broken for this run and no prompt-dependent delete is safe. For the other model, a quote you find verbatim in your own prompt covers the line there. A quote you cannot find is a different wording: ask that model , and only PRESENT covers the line. NONE, a quote that repeats the item, or a failed check leaves the line uncovered under that model. A quote covers the line only when the quoted sentence makes it a no-op, not when it shares the topic. Tool descriptions cannot be probed this way, because print mode loads fewer tools and is not among them; treat tool JSON as identical across models. A line the prompt covers under only one model is a keep.
