Opus 5 and Opus 5.5
Opus 5.5 has not been sampled; apply this file to it until it has.
Editing a draft this model wrote
Expect a clean lint report and a heavily staged draft. Do not let the report shorten the pass: spend it on steps 2b, 3 and 4.
- Read every header first. Rewrite each verdict into a label (entry 7). Expect most of them to be verdicts, some two sentences long: A cache doesn't make things faster. It makes some things faster and everything else slower.
- Read the last sentence of every paragraph next. If it restates the paragraph as a maxim, delete it (entry 12): Read-through caching is a p50 optimization sold as a latency optimization. Keep it only if it states a fact the paragraph has not.
- Split any sentence where
andorsojoins a fact to a generalisation of that fact (entry 39): …a second system that can be slow in new ways, and your users only ever feel the slow ways. - Cut significance designations the linter's noun list misses (entry 3): the real fix, the actual heuristic, that was the trap.
- Keep every technical claim. The argument under the staging is usually sound;
run
declaude_diff.pyto confirm the cuts removed none of it.
Running the pass as this model
- After rewriting several instances of one flagged shape, check whether the rewrites all landed on the same new shape. Verdict headers rewritten into nominalized headers (entry 41) is the usual one.
- Leave a title that states what happened in plain words. It is already a label. Turning Order dashboards undercounted a day's orders after a timezone change into dashboard undercount of one day's orders after a timezone change swaps a sentence for a noun pile (entry 41).
- Re-read every sentence you rewrote against entry 2. Rewrites of a staged line tend to land on a reversal: Finance found this problem; our monitoring never flagged it. State the fact once: No alert fired; finance found it.
- Read your rewrite once against entries 43 to 47 alone. Your clean prose
defaults to that register:
plainly,quietly,byte-identical,nothingstanding in for a search. - Lint the finished artifact, including your own annotations, summary and change list, not only the text you edited.
- If you also wrote the draft, run step 2b in a subagent or separate context. In PR #771 a pass by the drafting model let through six sentences the reader then flagged.
Evidence: oaustegard/experiments model-register-drift/ (two samples, one
prompt) and claude-skills PRs #769, #771, #779.