Wording tuned for older models in code-review, diagnosing-bugs, to-spec, ask-matt and wayfinder

[pihentagy]

I ran a prompt audit of these skills against Claude Opus 5.5 (Claude Code). I have not seen these fail in a real session yet, so this is an audit finding, not an observed failure. Current models follow the text below literally, which is why each line looks risky:

• code-review (both subagent briefs, SKILL.md:64 and :70): Under 400 words. The parent step aggregates the briefs verbatim, so on a large diff anything cut by the cap is lost.
  Proposed: Report findings only, one entry per finding, no preamble or restatement of the diff.
• diagnosing-bugs (SKILL.md:22): **Be aggressive. Be creative. Refuse to give up.** The Phase 1 completion criterion already sets the bar; the booster pushes past it.
  Proposed: Spend disproportionate effort here: the Phase 1 completion criterion below is the bar.
• to-spec (SKILL.md:33, :41): A LONG, numbered list… and extremely extensive. Read literally, this pads specs with near-duplicate stories.
  Proposed: A numbered list of user stories covering every actor and behaviour the feature touches, including edge and error cases. and drop line 41.
• ask-matt (SKILL.md:38, PHASE-BOUNDARIES.md:21) and wayfinder (SKILL.md:57): ~150k tokens vs one 100K token agent session. The numbers are tied to an unnamed model generation and disagree with each other.
  Proposed: drop the numbers ("the part of the window within which the model still reasons sharply", "one agent session").

The full change, with a patch changeset, is on a branch: main...pihentagy:skills:prompts-for-current-models

[pihentagy]

I experienced extremely long list of user stories for trivial features using /to-spec.

[lab1207]

I checked two of these against current main so you have a second pair of eyes before this merges:

Confirmed: to-spec padding. to-spec/SKILL.md says verbatim 'A LONG, numbered list of user stories' and 'should be extremely extensive and cover all aspects of the feature'. Read literally by current models, that is an instruction to pad — your replacement ('covering every actor and behaviour the feature touches, including edge and error cases') keeps the completeness intent while removing the length incentive. Take this one.

Qualify: diagnosing-bugs booster. The 'Be aggressive. Be creative. Refuse to give up.' line sits inside a tighter frame than the audit suggests: it follows the pass/fail-signal requirement ('If you don't have [a tight signal], no amount of staring at code will save you') and is immediately followed by five concrete signal-construction methods (failing test, curl script, CLI fixture diff, headless browser, trace replay). A model obeying the whole section allocates effort to signal construction, not unbounded flailing — the booster is bounded by its neighbors. I'd keep a milder version of your replacement ('the Phase 1 completion criterion below is the bar') but the current text is lower-risk than read in isolation.

Net: merge the to-spec change as proposed; soften (don't drop) the diagnosing-bugs line. Didn't check the word-count/token-number claims — leaving those to the branch reviewers.