Update agent-skills-creator for newer frontier model guidance

[mblode]

Summary

Folds the guidance from OpenAI's "Rethinking skills and prompts for GPT-6 Astra" into agent-skills-creator. Most of the post's points were already in the skill (situation-based triggers with the same migration example, grant-permission envelope, constraint cut, shared-repo model diversity). This PR adds the parts that were missing, without naming models or dates in the skill text.

Skills changed

• agent-skills-creator only

What changed

• Descriptions as a shared budget. A description is now judged in the installed listing next to its siblings, not alone: as short as it can be while still saying when it applies, because a host with too many descriptions trims all of them. New collection-workflow step in capability-delta.md audits every description as one listing before any body edit. Routing evals run against the whole listing.
• Review checkpoints as stopping points. "Stop after the first implementation and present for review" is now an explicit find in the constraint cut and audit dimension 8: keep it only where the decision is genuinely the user's, otherwise replace with a stated scope of done. New mode line, Step 3 wording, and a gotcha in SKILL.md.
• SKILL.md as a router for multi-workflow skills, in "Use the File System for Progressive Disclosure" and Step 3.
• Model-upgrade tension paragraph in capability-delta.md: cutting old rein-in constraints without adding scope of done and safe-loop permission yields a skill that stops sooner than before.
• Test-reminder example added to "Don't Instruct Behavior the Model Already Has", with the line between a generic reminder and a named completion check.
• Multi-vendor sourcing in "Test Across Models" replaces the Anthropic-only citation.
• Evals: two new cases (crowded-listing misrouting, skill that stops after the first draft) and two should-trigger prompts.

Validation

validate.sh skills/agent-skills-creator and validate.sh --all pass with 0 FAIL under a UTF-8 locale. With LANG unset the Ruby frontmatter parser fails for every skill in the repo, including untouched ones; that is a pre-existing environment issue in the validator, not touched here.

No README change: the skill count and bullet are unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YZaAGqf12Q9YHSm8vntmvj

---

Generated by Claude Code

---

Note

Low Risk
Documentation and eval scenario changes only for the agent-skills-creator skill; no runtime code or application behavior changes.

Overview
Updates agent-skills-creator guidance and evals for newer frontier-model behavior: crowded installed listings, literal review checkpoints, and collection-wide description routing.

Descriptions as a shared budget. Docs now treat every installed description as one listing the host may trim together—not per-file description-length passes. Collection workflow adds an audit-all-descriptions step before body edits; routing evals and rewrite steps emphasize shortening, moment-of-use triggers, sibling disambiguation, and testing against the full listing (bodies unchanged for routing fixes).

Completion vs early stop. Guidance flags generic “stop for review” / test reminders that frontier models follow literally, pairs constraint cuts with explicit scope-of-done and safe-loop permission on upgrades, and adds a SKILL.md mode plus gotchas for skills that hand back after a first draft.

Structure & evals. Multi-workflow skills should use SKILL.md as a router with per-workflow references. evals/evals.json adds scenarios for truncated-listing misrouting and post-upgrade early-stop behavior, plus matching should-trigger prompts.

^{Reviewed by Cursor Bugbot for commit 376cac2. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_03f7eed6-4f93-482e-b5c8-d28ffa13c137)