← ClaudeAtlas

preflightlisted

Challenges a draft before it goes out and hands back a revised version plus the facts only you can supply -- the six substance checks, then a set of reviewer personas whose wording problems it fixes. Never sends anything.
by-carlos/daikenja · ★ 0 · AI & Automation · score 63
Install: claude install-skill by-carlos/daikenja
# Preflight A bounded review loop. Run the substance checks, put the draft in front of reviewers who each read it for a different failure mode, apply what can be fixed, re-check once, and hand back a revised draft plus the questions only the user can answer. A real preflight is a walkaround where the pilot fixes what they find and looks again. Iteration is already in the name. **The rule that makes the loop safe, and the one this skill cannot break:** > **The loop may change wording. It may never change content.** Anything a reviewer raises that cannot be fixed from material already in the draft becomes a question back to the user. It never becomes an invented sentence. This skill has no send action. ## The model this context runs on Check this first, before Step 0. The safety hinge is Step 6, where every proposed fix is tested against the draft: one that needs an absent fact is reclassified instead of applied, and one that needs nothing outside the draft is applied instead of sent back as a question. That judgment is the whole difference between a loop that revises and a loop that invents, and it is materially better on Opus. **If you can tell you are not running on Opus, say this in one line, then carry on:** ``` Running on <model>. This loop's adjudication step is materially better on Opus -- `/model opus` and re-run if this message matters. ``` **This never blocks**, per `config-resolution.md`'s standing rule -- one notice line, then continue with reduced behav