Skip to content

docs(instruction-placement): tighten four skill bodies to the skill criteria - #11974

Merged
kyle-sexton merged 2 commits into
mainfrom
feat/6561-instruction-placement-concise-skills
Nov 1, 2026
Merged

kyle-sexton merged 2 commits into
mainfrom
feat/6561-instruction-placement-concise-skills

Conversation

@kyle-sexton

@kyle-sexton kyle-sexton commented Nov 1, 2026 •

Copy link
Copy Markdown
Member

Closes #6561

Summary

Rewrites four instruction-placement skill bodies to the skill criteria, wording only: audit, check, check-prerequisites and delta. The cuts remove intensifiers and restated emphasis ("perfectly", "exact", "wearing a different name", "It is not a bookkeeping detail"); no rule, threshold, identifier, cross-reference, frontmatter, code block or script changed.

Fix

Per-skill decision table:

Skill Ships Evidence
audit (SKILL.md + context/routing-out.md) rewrite 2 labelers PASS, 2 FALSE POSITIVE rows each
check rewrite 2 labelers PASS, 3 FALSE POSITIVE rows each
check-prerequisites rewrite 2 labelers PASS, 1 FALSE POSITIVE row each
delta rewrite 2 labelers PASS, 3 FALSE POSITIVE rows each
migrate, realign, setup unchanged they write files; the lane rule gives a state-mutating skill full before/after measurement even for wording-only edits, so they are left as they are

Unmeasured, and why: the plugin has no outcome eval cases, so the four rewritten skills ship on the two-labeler meaning diff alone, under the owner's Q7 decision for this rollout (recorded on #6476), which relaxes the AGENTS.md eval-evidence rule for wording-only cuts (issue Done-when allows naming them). Descriptions are byte-identical, so triggering is untouched.

Verification

  • Meaning diff: 8 docs-hygiene:compare-labeler runs, 2 per changed skill, before copy from git archive origin/main. Every row is FALSE POSITIVE; no SEMANTIC LOSS, AMBIGUITY or EMBEDDED INSTRUCTION row.
  • Pre-push check-affected passed on the rewrite push.
  • Merged origin/main: its one instruction-placement change since the base (docs(instruction-placement): align MS-13 cells with the re-pinned quote #11786: context/verified-mechanics.md, migrate/reference/verification.md) touches none of the rewritten skills, so the meaning diff stands.

Related

🤖 Generated with Claude Code

kyle-sexton and others added 2 commits November 1, 2026 00:13
…riteria

Wording-only pass over the audit, check, check-prerequisites and delta
skill bodies for #6561: drop intensifiers and one idiom, and lead the
check-prerequisites Next section with its successor. Frontmatter, fenced
blocks, headings, table rows, injection lines, CLAUDE_* tokens and link
targets are byte-identical; every rule, default, example and gate is
unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@kyle-sexton
kyle-sexton marked this pull request as ready for review November 1, 2026 16:07
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 1, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-01T16:08:43.827542Z dbaea1d Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dbaea1dd4b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

@github-actions

github-actions Bot commented Nov 1, 2026

Copy link
Copy Markdown
Contributor

claude-security-review reviewed dbaea1dd4bf3e173a2f58f961149ab132da1801a: read 6 of 6 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #11974

I found no security issues.

Scope: I read every hunk of all 6 changed files in .claude-lane/pr.diff:

  • .changes/instruction-placement/…-c44df31d.md, a new changelog fragment (patch bump)
  • plugins/instruction-placement/skills/audit/SKILL.md
  • plugins/instruction-placement/skills/audit/context/routing-out.md
  • plugins/instruction-placement/skills/check-prerequisites/SKILL.md
  • plugins/instruction-placement/skills/check/SKILL.md
  • plugins/instruction-placement/skills/delta/SKILL.md

Instruction-surface lens: The diff changes skill bodies, so I read it as a possible softening of standing instructions. Every hunk only cuts filler words or emphasis ("perfectly", "exact", "happened to", "wearing a different name", "It is not a bookkeeping detail") or reorders a sentence (the check-prerequisites Next line). Each rule keeps what it forbids or requires:

  • "Route, do not absorb" still bans dropping or reclassifying a routed candidate.
  • The suppression-surface rule is unchanged.
  • The delta hard rule that only the both-absent row routes out is unchanged.
  • The stale-line-range gotcha still says realign excises by that range.

No guardrail is removed or narrowed. The register file was not read. Recognition by consequence found no protected-class action that loses its check.

Other checks:

  • Trusted paths: None of the changed files match the TRUSTED PATHS: globs.
  • Trust controls: No trust control changes.
  • Plugin manifests: No manifest (plugin.json / marketplace.json) changes, so there are no manifest <field> lines.
  • Code: No workflow, script or other executable code changes.
  • PR context: The PR context has no embedded instructions aimed at the review lane.

critical-findings: 0

@github-actions

github-actions Bot commented Nov 1, 2026

Copy link
Copy Markdown
Contributor

claude-review reviewed dbaea1dd4bf3e173a2f58f961149ab132da1801a: read 6 of 6 changed file(s); 1 finding(s), 0 as inline comments.

Code review: PR #11974 (instruction-placement concise skills)

I read all six changed files in full: the changelog fragment, audit/SKILL.md, audit/context/routing-out.md, check/SKILL.md, check-prerequisites/SKILL.md and delta/SKILL.md.

Edits. Every hunk only changes wording. The edits remove intensifiers ("perfectly", "exact", "exactly"), shorten "somewhere else" to "elsewhere", drop the phrase "wearing a different name", and rephrase two sentences. No rule, identifier, cross-reference, frontmatter, code block or script changes. The reworded sentences still read as full sentences, and the reordered ## Next line in check-prerequisites follows the same skill-first form as migrate/SKILL.md. The bump: patch fragment fits a wording-only change. The path-scoped rules that match these files (skill-bodies-state-current-rules.md, cost-claims.md) are not broken: the edits add no history or cost claims.

Finding (blocking, closing issue #6561)

The issue's first Done-when criterion is not met: "The PR merges with a per-skill decision table in its body, and every shipped description and body was measured."

  • The PR body says the four rewritten bodies are "Unmeasured" and ship "on the two-labeler meaning diff alone".
  • The runbook the issue says decides what ships requires more than a meaning diff:
  • The protocol says to "ship only parts that were measured" and to re-measure output after any body change: skill-criteria.md#L219-L231.
  • The runbook's only basis for naming a skill unmeasured is the eval sandbox blocking a setup-style skill (#L78-L80). It is not a substitute for measuring a body that ships.
  • The PR cites a "wording-only rule" in the rollout. I found no such rule in the runbook or in skill-criteria.md. If it lives in Concise skills: one skill-creation standard, evals, and a fleet-wide rollout #6476, the PR should link it. The issue's second Done-when bullet ("Any skill left unmeasured … is named") reads as covering skills that were not rewritten, not as an exemption from the first bullet.

The issue's own thread says the same thing:

  • It lists "Approve the eval spend … before the measured passes" under "Needs a person".
  • The 2026-10-18 triage comment says this plugin's work runs in an attended eval session in batch 2 and is "not work for an unattended lane".

Next step. One of these resolves it:

A person needs to decide which applies.

Notes

  • Paved paths: these are wording-only edits, so they add no second way of doing anything and there is nothing to report.
  • Plugin manifests: no plugin.json or marketplace.json changed, so no manifest lines apply.
  • Verification: the meaning-diff results (8 compare-labeler runs, all FALSE POSITIVE) and the check-affected pass are the author's claims. I did not confirm them.
  • Embedded instructions: none found in the PR context.

@kyle-sexton

Copy link
Copy Markdown
Member Author

Reply to the claude-review summary on dbaea1dd4b:

# Finding Classification Evidence Reaction
1 #6561 Done-when "every shipped description and body was measured" not met; the "wording-only rule" is not linked INTENTIONAL The rule is the owner's Q7 decision for this rollout (2026-10-31T23:23Z, "go w/ recommended" to an option that stated it relaxes the AGENTS.md eval-evidence rule): wording-only rewrites ship on the two-labeler meaning diff; full measurement only for skills whose rules change. Now recorded verbatim on #6476 (#6476 (comment)) and linked from the PR body. The same rule shipped #6560 (#11938). It supersedes the Done-when measurement bullet for wording-only skills, and the four rewritten bodies are wording only (your review confirms "Every hunk only changes wording"). The eval-spend step is moot: no eval ran, and subscription usage is capacity, not cost (owner standing direction). 👀

@kyle-sexton
kyle-sexton added this pull request to the merge queue Nov 1, 2026
Merged via the queue into main with commit 1638d59 Nov 1, 2026
88 checks passed
@kyle-sexton
kyle-sexton deleted the feat/6561-instruction-placement-concise-skills branch November 1, 2026 16:24
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Concise skills rollout: instruction-placement (7 skills, 6 model-invoked)

1 participant