Repository navigation
feat(claude): tune skills, agents, and harness for Opus 5.5 / Sonnet 5.5 / Fable 5.1 - #252
Merged
Merged
Conversation
Opus 5.5 is the session model but consumers fell back to opus-5.md. Sonnet 5.5 backs the implementer/spec-reviewer/skeptic agents, so the README now routes agent-prompt tuning to the pinned model's profile. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The +7-10pt advisor gain was measured on the paragraph plus the hard rule, not the paragraph alone; the harness-supplied list is an observation, not documented; several quotes restored verbatim. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
implementer: stop at the briefed scope, require a check that exercises the change, and do not end the run on a plan or offer. spec-reviewer: report every finding with severity; the main session is the filter stage. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- workflow: consult the advisor before file-free design/risk answers - loop-engineering: a turn end is a report, not completion; cap automatic continuations at 2-3; wait for background work - settings: drop 'Think step-by-step.' from language (effort controls thinking; the field renders three times per prompt) - pentest: model-neutral safeguard note; audit with general-purpose, not Explore - issue-ship: cap the adversarial-review fix loop at 5 rounds Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- implementer: ambiguity no longer means stop at the start; finish the independent parts, then NEEDS_CONTEXT - workflow: advisor trigger is 'no file edits', not 'no tool calls'; drop the restated harness timing - loop-engineering: cap is 3 continuations per iteration, distinct from the review-fix cap; a hung background wait counts as stuck Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The language field rendered three times per prompt and was used as a prompt slot. Default behavior already follows the user's language; the CLAUDE.md line names the actual failure: drifting into English after English tool output. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r-e2e-test ask_user_input / web_search are claude.ai tool names and TodoWrite is not in this harness; the skills told the model to call tools that do not exist. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- ui-design-standards: defer to the repo's token source, name default looks to avoid, render each state, report all findings with severity; drop generic OOP/theme boilerplate - interrupt: agent commits only; push/PR wait for user confirmation; replace the unavailable wt tool with git worktree - ios-submission-review / zero-cost-scaling: drop stale statistics, verify drifting prices and rules before reporting - pentest-parallel-prs: drop duplicated recovery notes; stop condition finishes independent work first - descriptions: shorten pr-description and codex-subagent; drop a redundant loop-design flag; dedupe empirical-prompt-tuning Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- interrupt: dispatch implementer, branch from current HEAD, verify worktree placement, drop a nonexistent Agent parameter - trend-arbitrage: ask only for what is still unknown Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
skeptic is the highest judgment-per-token role with low output volume, so it gets the strongest evaluator; implementer and spec-reviewer stay on Sonnet. A fable subagent was confirmed to run through the business proxy. workflow.md's ~10K spawn estimate was far off: a one-line general-purpose probe used ~176K tokens, so narrow tools: lists matter. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a change is about to add a second unnamed layer, file, mechanism, or workaround, stop and decide: means out of proportion to the goal, or the goal itself needs questioning (the user's call). implementer reports instead of building disproportionate machinery; spec-reviewer flags over-engineering as a finding. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An agent-type Stop hook lists the user's typed and mid-turn requests from the transcript, checks the final reply against them, and sends one reminder naming anything neither done nor reported. It passes when stop_hook_active is set, judges coverage only, and defaults to allow when unsure. Simulated on this session's transcript with haiku: flagged the missing request in the positive case, allowed the complete reply in the negative case. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
…cisive evaluator Re-evaluated against the Opus 5 'no verification subagents' guidance: the cost half is adopted (narrowest agent, no self-started review rounds), the 'never' half is not, since separate agents found real defects on top of Opus 5.5's own checks in this session. Recorded as a hypothesis in the Opus 5.5 profile. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
概要
Opus 5.5 と Sonnet 5.5 の公式プロンプトガイドに合わせて、Claude 系のスキル、ルール、agents、設定を見直します。Fable 5.1 については 2026-09 に作ったプロファイルを参照しています。既存のハーネスは Opus 5 / Fable 5.1 世代向けに整っていたので、全面的な書き直しはしていません。ガイドが具体的に指摘している差分と、読み合わせで見つかった不整合だけを直しました。
差分は +184 / −182 行で、すべてプロンプト・設定・ドキュメントです。中心となる変更は「主な変更」の前半 3 項目です。残りの各スキルの直しは小さな修正の積み重ねです。
主な変更
モデルプロファイルを追加した（
references/model-profiles/）opus-5-5.md: Opus 5 からの差分。effort の既定値がmediumに下がったこと、thinking を無効化できないこと、無人実行中にツール呼び出しなしでターンを終える早期停止パターン、advisor の測定範囲などを書いたsonnet-5-5.md: implementer と spec-reviewer が sonnet で動くため追加した。agent を調整するときは、その agent が使うモデルのプロファイルを読むよう README に書き足したagents
model: fableに変えた。判断が重い一方で出力量が少なく、Fable の単価が効いてくる箇所が限られる。業務 proxy 経由でも fable のサブエージェントが動くことを確認したrules / settings / CLAUDE.md
development-principles.md: 「複雑さは停止の合図」を追加した。依頼が名指ししていない仕組みを 2 つ目に足す前に止まり、手段が過剰なのか、目的を問い直すべきなのかを判断するworkflow.md: ファイルを編集しない設計判断でも advisor を呼ぶ。spawn の固定費が実測で約 10K を大きく超えていたので、ツールを絞った agent を選ぶよう書き換えたloop-engineering.md: ターンの終了は報告であって完了ではない。自動継続は 1 iteration あたり 3 回までとし、バックグラウンドの処理が終わるのを待つsettings.json.tmpllanguageフィールドを削除した。値がシステムプロンプト内で 3 回展開され、プロンプトの代わりに使われていたためCLAUDE.mdの 1 行で補正するsettings.json.tmplに agent 型の Stop hook を追加したスキル
ask_user_input/web_search、browser-e2e-test のTodoWrite、interrupt のwtgenerator-evaluator 分離を見直した（
harness-engineering.md）。Opus 5 ガイドの「検証用のサブエージェントは不要」には 2 つの側面がある意図して変えなかったもの
検証
just test→✓ All checks passed!（最終コミット時点）fce8f8f）🤖 Generated with Claude Code