Skip to content

Catch up Tier-A skills (review/flywheel/ship/plan/codify) to June 2026 - #4

Open
mrskiro wants to merge 2 commits into
mainfrom
update/tier-a-skills-authority-catchup
Open

Catch up Tier-A skills (review/flywheel/ship/plan/codify) to June 2026#4
mrskiro wants to merge 2 commits into
mainfrom
update/tier-a-skills-authority-catchup

Conversation

@mrskiro

@mrskiro mrskiro commented Jun 22, 2026

Copy link
Copy Markdown
Owner

Summary

PR #3 で agentic-audit と skill-creator は完了済み。本PRは Tier-A 5 skill (review / flywheel / ship / plan / codify) に同じ authority delta を適用した第2陣。

review

  • Osmani の 93%-by-one heterogeneity finding を「複数reviewerを走らせるなら違うcharacterを選べ」として導入
  • "First human to lay eyes on this code" 問題 → PR本文にdecision logを残す
  • Don't use Haiku as strict-domain verifier (48% false-pass / LangChain × Harvey)
  • Batch verification 10× cheaper than per-criterion
  • Reviewer-prompt の "find only correctness gaps" 注意喚起

flywheel

  • LangChain 4-loop の Loop 4 hill-climbing として位置付け
  • 2026 OTEL surface 拡張 (skill_activated.invocation_trigger / skill.kind=workflow / plugin events / ToolSearch fetches / agent_id+parent_agent_id / hook_execution_complete 構造化 等)
  • Corrections-become-bounded-evals pattern (OpenAI tax + LangSmith Engine)
  • "What can I stop doing?" 削除判断ロジック

ship

  • Step 3.5 Runtime smoke ("Step 8" Shogun): build green ≠ DB up, test green ≠ dev server up
  • Anti-rationalization table (Osmani)
  • Auto mode -p abort behavior 警告

plan

  • Interrogatory LLM (Harper Reed / Fowler) — one question at a time with recommended answer
  • Strategic (user) vs tactical (agent) split (Karpathy)
  • 4 conversation registers (Chelsea Troy) — plan = "Deciding"
  • SPEC.md frozen artifact pattern (Anthropic best practices)

codify

  • Skill-internal codify target (Gotchas / anti-rationalization / ${CLAUDE_PLUGIN_DATA})
  • VibeSec security context file
  • Step 7 "What can I stop doing?"
  • Cognitive debt warning (Joshi)

Diff

+125 / -13 across 5 files:

  • dot_claude/skills/review/SKILL.md (+24/-3)
  • dot_claude/skills/flywheel/SKILL.md (+22/-2)
  • dot_claude/skills/ship/SKILL.md (+29/-0)
  • dot_claude/skills/plan/SKILL.md (+31/-3)
  • dot_claude/skills/codify/SKILL.md (+32/-5)

Test plan

  • Trigger /review with a real diff — verify Codex CLI is still called and Codex output is presented verbatim
  • Trigger /plan — verify the agent only asks strategic questions one at a time with a recommended answer
  • Trigger /ship with a UI change — verify Step 3.5 runtime smoke is exercised
  • Trigger /codify after fixing a skill-related bug — verify the Gotchas-first / anti-rationalization-first priority is followed before docs/
  • Run /flywheel and verify new OTEL fields are queried (or the existing queries still pass with the older schema)

🤖 Generated with Claude Code

mrskiro and others added 2 commits June 22, 2026 10:24
@mrskiro
react: add render purity rule (no ref.current writes during render)
13e9ee9
@mrskiro @claude
chore(claude): catch up Tier-A skills (review/flywheel/ship/plan/codi…
1a2a766 
…fy) to June 2026

April→June の authority output を反映した第2陣 (#3 で agentic-audit と
skill-creator は完了済み)。

review:
- Osmani の 93%-by-one heterogeneity finding (4 reviewers / 146 PRs / 679 findings)
- "First human to lay eyes on this code" 問題 → PR本文に decision log を残す要請
- Don't use Haiku as strict-domain verifier (48% false-pass)
- Batch verification 10× cheaper than per-criterion
- "Reviewer prompted to find gaps will find some" 警告

flywheel:
- LangChain 4-loop taxonomy 上での位置付け明示 (Loop 4 hill-climbing)
- 2026 OTEL surface 拡張: mcp_server.name × mcp_tool.name 分離, skill_activated +
  invocation_trigger, skill.kind="workflow", plugin/marketplace events,
  ToolSearch fetches, agent_id/parent_agent_id, hook_execution_complete 構造化,
  at_mention / permission_mode_changed / api_refusal
- Corrections-become-bounded-evals pattern (OpenAI tax + LangSmith Engine)
- "What can I stop doing?" 削除判断 (Anthropic Harnessing Claude's Intelligence)

ship:
- Step 3.5 Runtime smoke ("Step 8") — build green ≠ DB up, test green ≠ dev server up
- Anti-rationalization table (Osmani): tests-pass-ship-it / I'll-test-later / etc.
- Evidence over claim principle
- Capture intent for the reviewer when dispatching worktree subagents
- Auto mode -p abort behavior (3 consecutive / 20 total denials → terminate)

plan:
- Interrogatory LLM pattern explicitly cited (Harper Reed via Fowler)
- Henry v1.3.0 / Anthropic-best-practices: recommend an answer + AskUserQuestion
- Strategic (user) vs tactical (agent) decision split (Karpathy)
- "Don't push user with A/B/C/D" anti-pattern (Shogun)
- Chelsea Troy 4 conversation registers: plan = "Deciding" register
- SPEC.md frozen artifact → fresh session for /ship

codify:
- Codify target hierarchy expanded: skill-internal (Gotchas / anti-rationalization /
  ${CLAUDE_PLUGIN_DATA}) > security context file > docs/ > rules/ > CLAUDE.md
- Anti-rationalization table addition pattern (Osmani)
- VibeSec security context file as codify target for stakes-bearing rules
- Step 7 "What can I stop doing?" — codify is also about removing
- Cognitive debt warning when codifying introduces new vocabulary (Joshi)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@mrskiro