feat(claude-config): add the unhobble skill and codify instruction economy - #2005
Conversation
…onomy
Digest-driven alignment with Boris Cherny's YC Startup School talk ("We Cut
80% of Claude Code's Prompt") and the official best-practices doctrine it
matches. Two deliverables:
- docs/PLUGIN-PHILOSOPHY.md gains an "Instruction economy" section: standing
instructions are a per-session tax; additions are evidence-gated (repeated
observed stumbles, never anticipation); ablation is generation-triggered;
evals outlive instructions; deterministic policy hooks and team conventions
are the officially carved-out durable tier; model-capability claims never
relax the security posture.
- New claude-config:unhobble skill (0.22.0): the empirical bare-baseline
experiment — snapshot/classify, reversibly strip behavioral surfaces on a
dedicated branch, observe real stumbles in fresh sessions via a ledger,
re-add only what two-plus same-cause rows earn back, each restore citing
its evidence. audit-instructions gains the reciprocal route-out.
Verified by two fresh-context reviewers (skill-quality contract + research
fidelity); check-skill.sh PASS, markdownlint clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
|
Claude finished @kyle-sexton's task in 3m 44s —— View job Code review complete
Scope: the 7 files changed in this PR ( Findings (posted inline)
Otherwise
|
|
Claude finished @kyle-sexton's task in 1m 14s —— View job Security review complete ✅
Scope: all 7 files changed in this PR (commit Findings: none. This PR has no security-relevant surface to flag:
No CRITICAL, IMPORTANT, or SUGGESTION findings to report. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3185e23c6b
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".
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
Claude has reviewed this PR 5 times. The lane skips further automatic reviews after 5; deleting this comment resets the count. |
Main released its own 0.22.0 (#2004) while this branch was in flight; the unhobble release moves to 0.23.0 with both changelog sections preserved. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 38s —— View job Security review complete
Scope: all 7 files changed in this PR ( Findings1. [IMPORTANT] Plugin-level strip granularity can silently disable a policy hook the skill itself promises to protect — confidence: high
This repo has a concrete instance of exactly that mixed plugin: Recommend extending Phase 1's classification to per-hook granularity inside each project-enabled plugin (not just project-settings hooks), so a mixed plugin's policy hooks are individually preserved even when its behavioral skills are stripped. 2. [IMPORTANT] Experiment state keyed by repo basename + model + date collides across checkouts sharing a basename — confidence: high (independently confirmed, not just plausible) Already flagged by @chatgpt-codex-connector on Otherwise
|
|
Claude finished @kyle-sexton's task in 1m 33s —— View job Code review (re-run after synchronize)
What changed since the last reviewThis synchronize is a merge of Sanity-checked the merge itself:
Prior findings — still open (unaddressed by this synchronize)Since the reviewed content didn't change, the two substantive findings from the earlier pass on
OtherwiseNo new issues introduced by this synchronize. Version/changelog/description bump to 0.23.0 is |
…ption Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 48s —— View job Security review complete
Scope: Findings: none.
No CRITICAL, IMPORTANT, or SUGGESTION findings to report for this synchronize. |
|
Claude finished @kyle-sexton's task in 48s —— View job Code review (re-run after synchronize)
What changed since the last reviewThis synchronize adds one commit, Checked that this regeneration is correct and mechanical, not hand-authored:
No new findings from this delta. Prior findings — still open (unaddressed by this synchronize)Content in
OtherwiseNo new issues introduced by this synchronize. The |
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 33s —— View job Security review complete
Scope: this synchronize adds exactly one commit, Findings: none.
No CRITICAL, IMPORTANT, or SUGGESTION findings to report for this synchronize. |
|
Claude finished @kyle-sexton's task in 40s —— View job Code review (re-run after synchronize)
What changed since the last review
Checked this is mechanical, not hand-authored content requiring independent review:
Every other file in the PR ( Prior findings — still open (unaddressed by this synchronize)Content in
OtherwiseNo new issues introduced by this synchronize. The |
Three review findings on #2005: - Experiment state now keys on canonical checkout identity: the id gains a run nonce, the manifest records the resolved worktree path and origin URL, later phases verify the match and abort on mismatch, and snapshot never reuses an existing experiment directory — two same-basename checkouts can no longer cross-restore each other's settings. - Phase 1 classification now covers every surface Phase 2 acts on (instruction files and project-enabled plugins included), and mixed instruction files are split at section granularity so the convention carve-out survives a strip instead of being deleted with the file. - The 80%-cut citation in PLUGIN-PHILOSOPHY.md is now a clickable URL to the Anthropic blog post, matching the neighboring best-practices citation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 37s —— View job Code review (re-run after synchronize)
What changed since the last reviewNew commit
New findings (posted inline)
Otherwise
|
|
Claude finished @kyle-sexton's task in 1m 20s —— View job Security review complete
Scope: this synchronize ( Prior findings — resolution status1. Experiment-ID collision (previously IMPORTANT, confidence: high) — now fixed. 2. Mixed-plugin policy-hook bypass (previously IMPORTANT, confidence: high) — still open. But the same section still lists "a plugin" as a single atomic classification unit alongside "a hook entry, a rule file, a skill, an agent" (line 86) — it does not extend the mixed/split treatment to a plugin's internal components. Phase 2's plugin handling (L105-106) is unchanged from the version already flagged: "disable the ones classified Concretely, this repo's own Otherwise
Summary: 1 IMPORTANT finding remains open (mixed-plugin policy-hook bypass); the previously-open experiment-ID collision finding is resolved by this commit. |
…with evals Second review round on #2005: - A plugin bundling any policy-classified surface is kept whole (plugins toggle whole; no partial disable), its behavioral components recorded as unstripped-mixed-plugin and flagged as observe-phase confounds — the policy gate never goes down with the plugin. - Three new evals lock the recent rails: mixed-plugin-kept-whole, checkout-identity-mismatch-aborts, mixed-claude-md-split-not-deleted-whole. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 1m 21s —— View job Security review complete
Scope: since the last security-lane review ( Prior finding — now resolvedMixed-plugin policy-hook bypass (previously IMPORTANT, confidence: high) — fixed. Commit This also picks up the companion nit from the last pass: Findings: none. No new security-relevant content was introduced by this synchronize, and the previously open finding is resolved as described. No workflow files, hooks, or executable scripts are touched by this PR at any point — every file remains Markdown/prose, a manifest version/description bump, or a natural-language Summary: 0 open findings. Both IMPORTANT findings raised across this PR's review history (experiment-ID collision, mixed-plugin policy-hook bypass) are now resolved. |
Third integration. Main added claude-config's unhobble skill (#2005), which rewrote the manifest description to name six skills. Resolution keeps this branch's version (0.23.0, the higher bump) and main's description, since mine carried only the stale five-skill text. Both 0.23.0 changelog entries sit under the one Added heading — they describe different additions to the same release. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NNmc2T3y81efZTaDZHerWo
…alignment Main's #2000/#2005 landed their own I23 (context-budget, fable-5), catalog 1.17.0, claude-config 0.23.0, session-flow 0.19.0, work-items 0.33.0 — colliding with this branch's identically-numbered additions. Resolution: - claude-config 0.24.0, catalog 1.18.0; this branch's rows renumbered I23/I24/I25 -> I24/I25/I26 with cross-references updated; partition and SKILL.md ranges to I26; Sources merged (Opus 4.8 guide + whats-new-sonnet-5 added alongside main's entries); I8-b/I8-d/I8-e edits re-applied onto main's revisions of those rows. - session-flow 0.20.0, work-items 0.34.0; both CHANGELOG entries reinserted above main's same-numbered releases. - New rows checked against main's ADR 0008 (present-text anchoring): all four anchor to present passages. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HTJ15VxULScgiYFWFRSV8G
…2015) No linked issue ## Summary Aligns this repo — and its consumer-facing audit machinery — with the [Sonnet 5 prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-sonnet-5). The guide was read in full from the raw-`.md` channel (MD5 `6d23959f0ed226feb06bf20c314029e3`, byte-identical to the repo's recorded 2026-08-04 snapshot), broken down into a 13-concern inventory, and applied in both directions the guide binds: this repo's own instruction surfaces, and the `audit-instructions` catalog downstream consumers run against their repos (`/claude-config:audit-instructions --target-model sonnet-5`). ## Fix **Consumer-facing catalog (`claude-config` 0.24.0, catalog 1.18.0)** — every row verified 2026-08-08 against raw-`.md` fetches with recorded byte sizes + MD5s (rows renumbered after merging main's Fable 5 alignment, which took I23 and catalog 1.17.0): - New **I24** instruction relying on silent generalization (unscoped — Sonnet 5 and Opus 4.8 guides state the literalism claim verbatim-identically). - New **I25** sampling parameter prescribed where the model rejects it (unscoped; Detect range Opus 4.7+, Sonnet 5, Fable 5, Mythos 5; 400-class `error` like I17/I20). - New **I26** generic negative steering on open-ended design briefs (unscoped — both guides converge; propose-options is the documented variety mechanism now that `temperature` is rejected). - New **I17-d** tool reliance with thinking disabled and no explicit tool nudge (`sonnet-5`-scoped; the coupling is stated only in that guide — Opus 4.8's differing default recorded as the scope negative). - **I8-e unscoped**: its own recheck trigger fired — the Opus 4.8 guide now states the forced-interim-status-cadence claim, meeting the promotion gate. I8-d cedes the cadence shape (one finding per line). I8-b corroboration extended; Sources gain the Opus 4.8 guide and whats-new-sonnet-5. **Repo self-alignment** (guide § "Code review harnesses", "More literal instruction following", "Design and frontend defaults"): - `review` 0.17.0 — fanout gains a dispatch contract (finder leaves told coverage is the job; the 5-stage normalization pipeline stays the filter); per-slice template reports coverage-first with a Confidence column; three agents now enumerate **every** `REVIEW.md` code-span citation instead of "a citation"; security-reviewer gains a stated floor for unlisted ecosystems; ecosystem-specialist reports UNVERIFIED instead of silently skipping ecosystems with no bundled default; untracked-files read line added to the two agents missing it. - `session-flow` 0.20.0 — retro quick mode's findings bar is a decidable test, not "significant"/"minor". - `work-items` 0.34.0 — `work` Step 4 staleness pre-check checks every concrete referent; bullets labeled as examples. - `prototype` 0.4.0 — explore-directions makes the visual axis additive: mockup variants declare background/accent/typeface + rationale; structure stays the floor; eval 3 no longer treats color variety as out of scope. Audit lanes that came back clean (no drift found): sampling params, thinking-disable prescriptions, cadence scaffolding, token budgets, effort frontmatter (all `high` = model default), stale Sonnet model facts (severity.md already cites and follows the guide), design steering outside prototype, repo-root surfaces. ## Test plan - `markdownlint-cli2` clean over every edited markdown file. - `jq` validation of all bumped `plugin.json` files and the edited `evals.json`. - Fresh-context verification agents: implicit-scope and design-variety audit findings verified against files before fixes were applied; an independent branch-diff verifier script-checked quote fidelity of every upstream citation against the raw-`.md` snapshots across three passes (16 defects raised, all fixed; final verdict CLEAN). The Opus 5 and Fable 5 guides were additionally fetched to close the verifier's disclosed snapshot gaps. - Merge with main's Fable 5 alignment (#2000) and unhobble (#2005) PRs resolved semantically: rows renumbered around main's I23, versions advanced past main's, re-applied edits re-linted; new rows checked against main's ADR 0008 present-text admission rule. - CI: pr-title, pr-issue-linkage, and repo lint workflows on this PR. ## Verification Guide + adjacent pages (whats-new-sonnet-5, migration guide § Sonnet 5, effort, thinking § effort, Opus 4.8 guide) fetched via the raw-`.md` channel on 2026-08-08; all quoted spans in catalog rows carry verification stamps with byte sizes + MD5s. The Sonnet 5 guide is byte-identical to the 2026-07-29/2026-08-04 captures recorded in `playbooks/reference/model-adaptation/sonnet-5.md`, which therefore needed no delta. ## Related - `plugins/playbooks/reference/model-adaptation/sonnet-5.md` — session-arming counterpart (already current; hash-verified this session). - Prior guide-driven work: review 0.16.0/0.16.1 severity.md decidable-test rewrite (already aligned). 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01HTJ15VxULScgiYFWFRSV8G --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…#2020) No linked issue ## Summary Adds the first row to `.github/recurring-schedule.json`, operationalizing the instruction-economy doctrine's generation-triggered ablation (merged in #2005): on each frontier model release, run `/claude-config:audit-instructions` + `/claude-config:unhobble` over this repository's session surfaces. Until now the practice existed only as doctrine prose — nothing prompted it when a release actually lands. ## Fix One schedule row, `frontier-release-instruction-ablation`: primary trigger `frontier-model-release` (the release event, not the date, per docs/PLUGIN-PHILOSOPHY.md "Instruction economy"), quarterly cadence as backstop only, `next_due` 2026-11-06. Scope note pins the pass to session surfaces, never the shipped plugin components. Backed by tracker issue #2019 (`[Maintenance]`-prefixed, `recurring` label) so `/work-items:track due`/`recheck` and the `/work-items:work` due-recurring tier can see and reconcile it. ## Verification - `jq . .github/recurring-schedule.json` parses clean; row matches the schema used by the work-items `add`/`recheck` actions (id, cadence, triggers, last_checked, next_due, close_previous). - Issue #2019 created through the work-item-tracker seam with exact `[Maintenance] {title}` match to the schedule row title, as `recheck` requires for close-matching. ## Related - Refs #2019 — the recurring tracker issue this schedule row backs (not closed; it is the recurring instance) - Refs #2005 — the merged instruction-economy doctrine + unhobble skill this cadence operationalizes 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…, alias-abort UX, chapter discoverability) (#2017) Closes #1996 ## Summary Executes the interview outcomes over #1996's six deferred Opus 5 design candidates: decisions (b) effort-misconception audit row, (d) opus-5 chapter discoverability, and (e) family-alias abort UX are accepted and shipped here; (a) generic prior-model-workaround row and (c) vision-workaround row stay deferred-with-trigger; (f) the additive chapter-payload omissions are ratified as decided-not-missed. All six outcomes are recorded in a comment on the issue. ## Fix - **claude-config 0.25.0** (criteria 1.18.0 → 1.19.0): - New catalog row **I27 — effort lowered to shorten the response** (renumbered twice from I23 as main's #2000 and #2015 landed I23–I26 concurrently), `Model scope: opus-5` (both source statements are model-qualified; promotion gate unmet, unscope trigger on the row), with cost/latency-ground, length-instruction-only, audience-test, and config-value fences. - New `instruction-scan.sh` **I27 family**: effort-lowering directive + brevity token ANDed on one line, word-boundary conventions applied, doubled-consonant `dropp?` stem per review; header/usage docs updated; 11 I27 regression assertions (57 checks total). - **Family-alias abort UX**: the fail-loud abort on a version-ambiguous target now also names the normalized token the alias currently resolves to as a suggested `--target-model` value the user confirms — still never proceeds on a guess. - Stale `evals/evals.json` check-range fixed (I6-I16 → I6-I27; `--opinion` set corrected to I16, I19, I22). - **playbooks 0.7.1**: `fable-5` description now names the per-model adaptation chapters (Opus 5, Opus 4.8, Sonnet 5) with Opus 5 trigger phrases — the opus-5 chapter was previously unreachable from the skill listing. ## Verification - Both guide quotes verified against the LIVE Opus 5 prompting guide this session: raw `.md` 11,225 bytes, MD5 `8579d63fc9f793784b8c56320fd74e71`, byte-identical to the 2026-07-25 knowledge-corpus capture; the effort page's Opus 5 section corroborates the claim unhedged. A sibling fresh-context verifier independently re-fetched the page and reproduced every quote and both hashes. - `instruction-scan.test.sh`: 57/57 checks pass. - `check-changelog-parity.sh --check`, `--check-bump origin/main`, `--check-order`: all pass after each rebase (branch rebased over #2005, #2000, #2015; versions re-slotted each time). - `check-changed-skills.sh origin/main`: 2 skills checked, 0 failed; fable-5 description 828 chars (cap 1,536); aggregate listing budget 1,801/8,000. - Fresh-context reviewer pass applied: prose-grammar fix in the seeding list, word-boundary consistency on the brevity ERE, brevity-token test-matrix completion, doubled-p stem fix, evals range fix. ## Related - Refs #1699 (the workstream that shipped the built Opus 5 audit set) - Refs #1997 (the repo-alignment PR; carried the chapter half of decision c) - ADR-0006 / ADR-0007 (model-scoping promotion gate; doctrine hosting seam) 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01E5qiaLiGqP6fUTHkRYS19R --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…guide and add the posture-audit lane (#2018) ## Summary Aligns the marketplace's prompt-bearing surfaces with the current official prompting guide (best-practices page + the Fable 5 / Sonnet 5 / Opus 5 / Opus 4.8 subpages, all fetched 2026-08-08), and ships the tooling downstream consumers need to run the same alignment on their own repos. Rebased over — and reconciled with — the parallel alignment PRs #2000 (Fable 5) and #2015 (Sonnet 5): where those landed the same catalog change first (Opus 4.8 in Sources, I8-e unscoping, a sampling-parameter row), this PR keeps main's version and contributes only what remained novel. **Repo alignment (audit → adversarial verification → fix):** every plugin's instruction surfaces were audited against a 30-item checklist distilled from a verified 451-item inventory of the guide; every finding was re-adjudicated by independent fresh-context verifiers before any edit. ~55 verified fixes across 21 plugins (each with a version bump and CHANGELOG entry), including: - machine-health: three high-severity internal contradictions — two reference files still described the pre-`approvals.json` approval model (checkbox-driven TODO.md, writes into the plugin install directory); plus a broken markdown fence-nesting bug in the report template and a wrong-schema pointer in setup - songwriting: "Pat's seven title types" stated three mutually inconsistent taxonomies across four files — reconciled to `hook.md`'s canonical list - playbooks (boris): stale model/effort rows (`/model` → `/effort`, missing `xhigh`, superseded model recommendations) corrected per the pack's own supersession-note conventions, verified against live docs - implementation + testing: the repo's two code-editing loops (implement's TDD cadence, diagnose's fix step) gained anti-test-gaming guards (guide: "Avoid focusing on passing tests and hardcoding") - ~30 instruction-style fixes (emphasis-language, missing rationale, example/structure hygiene, scope explicitness) and removals of forced status-cadence and fabricated context-percentage triggers **Downstream capability:** - `claude-config:audit-instructions` criteria 1.18.0 → 1.19.0: new row **I27** (over-aggressive trigger emphasis / blanket tool defaults — the one guide concern no parallel PR had covered); **I8 base promoted to unscoped** via the model-agnostic prefer-general-instructions statement, with the delegation-throttle worked instance kept `fable-5`-scoped (the Opus guides recommend the opposite shape); **I21** gains the calibration-staleness-vs-level-adequacy distinction. `instruction-scan.sh` now seeds I27 (both arms) and the existing I25 sampling-parameter row. - New skill `claude-config:audit-prompting-postures` — the additive lane: detects posture guidance (delegation criteria, minimal-scope/anti-test-gaming guardrails, grounding, autonomy/destructive postures, etc.) that the guide says a component's purpose needs but the component does not carry; report-only, applicability-predicated, wording fetched live (pointer-not-copy) - `docs/OFFICIAL-DOCS.md`: new "Prompting doctrine (platform docs)" section indexing the guide pages + Thinking page, composed with #2014's evaluation-guidance section **Follow-ups worked to completion in this PR** (originally filed as issues, then implemented here on approval instead of being left parked): - Closes #2026 — `claude-config:audit` Phase 1 now reads machine-scope managed settings (OS-resolved path + `managed-settings.d/` drop-in, structure-only, no-secrets posture); the checklist tick is restored, now truthful - Closes #2027 — `audit-pass`'s 892-line `run-contract.md` split into five per-topic reference files behind a routing index, §-numbering preserved, inbound links repointed, last history-narration clause reframed - Closes #2028 — `course-digest`'s two inconsistent self-estimated context-percentage handoff triggers replaced with one qualitative signal set (long/degraded session, compaction, module boundary); resume protocol untouched - #2029 closed separately as already-covered: `lib/hook-utils.sh` is the canonical source and CI's `hook-utils-sync` required check already gates every plugin copy Also swept: the `rhyme` skill's latent skill-relative `context/pat-pattison/` mention (same defect the CI gate caught on its seven edited siblings) is now plugin-root-relative. ## Related - Prompting best practices: <https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices> - Model subpages: prompting-claude-fable-5, prompting-claude-sonnet-5, prompting-claude-opus-5, prompting-claude-opus-4-8 (same section) - Thinking page (sampling-param 400 set, disable-rejection set): <https://platform.claude.com/docs/en/build-with-claude/thinking> - Parallel alignment PRs reconciled with: #2000, #2015, #2005, #2014 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01LYm1iCn68WiKKsWbP994Co --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
1 participant
Summary
Alignment change set from the full
/knowledge:youtube-digestof Boris Cherny: "We Cut 80% of Claude Code's Prompt" (YC Startup School, 2026-07-25). Research record + prioritized menu live inmelodic-software/knowledge-corpusPR #10 (sliceboris-cherny-we-cut-80-of-claude-code-s-qyPCVqFUyDo).Repo alignment (menu P0 items):
docs/PLUGIN-PHILOSOPHY.md— new Instruction economy section: per-session-tax framing, evidence-gated additions ("stumble twice" before any new standing instruction), generation-triggered ablation, evals-outlive-instructions, the official durable-tier carve-outs (deterministic policy hooks, team conventions in git), explicit security non-relaxation (injection-resistance claims are hedged in primary sources), and the note that verification-first task design is already encoded by theverification/planning/tdd/testingplugins.Downstream-consumer capability (menu #2):
claude-config:unhobbleskill (0.22.0) — the empirical bare-baseline experiment consumers run on their own repos: snapshot + policy-vs-behavioral classification → reversible strip on a dedicated branch → stumble ledger across fresh sessions → evidence-gated re-add citing ledger rows. Managed settings and policy hooks are never stripped;CLAUDE_CODE_SIMPLEexplicitly out of contract (undocumented). 8 evals;check-skill.shPASS (0 errors, 0 warnings, description 641/1536);audit-instructionsgains the reciprocal route-out.Verification: two fresh-context reviewer agents (skill-quality contract; research-fidelity vs the slice's RESEARCH.md) — all findings fixed (skill count phrasing, blog citation + "coding evaluations" qualifier, reciprocal routing, description trim).
Deliberately deferred (tracked in the slice's
recommendations/): skill-authoring playbook echoes of the evidence gate (avoids doc duplication — the instruction-economy rule itself argues against second copies), listing-budget program (menu #4, needs operator decision), hook-surface policy-vs-behavioral classification sweep (menu #5), marketplace self-maintenance routines (menu #9), recurring work-item wiring for the ablation cadence (menu #1 second half).No linked issue
Related
🤖 Generated with Claude Code