All skills
aktsmm avatar

/opportunity-factory

@e63d714
by yamapanaktsmm/agent-skills26 stars
4

Run a reusable opportunity-to-artifact workflow: discover unmet needs, set up workspace factories, schedule recurring commander/worker/reporter prompts, batch-refine many items, optionally use SQLite state, build small artifacts, review quality, and track outcomes. Use when the user wants to repeatedly create apps, games, products, content, or experiments from market/user needs.

Use this Skill: https://skilld.dev/gh/aktsmm/agent-skills/opportunity-factory

This session only. Nothing lands on disk.

referencestunable-defaults.md

≈4.5k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Tunable Defaults

Skill が示す数値・順序・cadence・rubric 判定基準・質問セット等は reference default。AI が実運用で改善して良い。Hard rule (下記) だけは変えない。

変更主体と Autonomy Mode

  • 変更主体は workflow-review role のみ。worker / commander / reporter-learner は変更不可 (change propose も禁止、observation の report のみ可)
  • Autonomy Mode に応じて propose or apply が決まる。SSOT は references/runtime-modes.md の ai-autonomous preset 内 "Tune Apply by Autonomy Mode" table

概略:

  • Normal / AUTO: workflow-review が propose、user 承認で apply
  • FULL / ALL: workflow-review が propose、自動 apply (hard rule 誤変更疑いは security-approve escalate)
  • 3 サイクル追跡 revert は reporter-learner が担当 (全 mode 共通)

Workflow-review Dispatch

Weekly cadence + 以下 ad-hoc trigger で hourly cycle 内でも dispatch (fallback lane #3 の advisory critic とは別レイヤー):

  • Blocker Test 4/4 escalation が 24h で 3 件以上
  • Anti-pattern registry の同一 fingerprint count が K=3 到達
  • Discovery Floor trigger が 3 サイクル連続でも新規 candidate 流入ゼロ
  • Critic-log の Layer 3 reject が 24h で 2 件以上

Ad-hoc trigger 条件も tunable、実運用で調整可。

変更可能項目一覧

項目 既定値 変更判断シグナル 変更手順
Fallback lane 順序 品質重視 A (1..10) Discovery 偏り、lane 特定枯渇 workflow-review propose → smoke test → apply
Discovery Floor cycles 5 Portfolio 停滞率、Discovery 偏り workflow-review 調整
Persistence Standard max 12 / stall 4 / replan 4 / approach 4 / dim 2 Diminishing returns 増加 task class 別に調整可
Persistence Persistent max 20 / stall 5 / replan 5 / approach 6 / dim 3 同上 同上
Persistence Exhaustive max 30 / stall 6 / replan 6 / approach 8 / dim 4 同上 同上
Review repair workflow rounds 10(下限 3 / 上限 20)。既存 persistence の残り iteration と小さい方を上限にする unresolved finding、replan率 workflow-review が範囲内で調整
Independence block limit 3(下限 1 / 上限 5)。同一 parent task の独立critic不在を parked-independence へ上げる回数 adapter/model outage、approval数 workflow-review が範囲内で調整
Task class → Profile mapping portfolio/cleanup/digest=Standard, worker本流/discovery/small-bet=Persistent, prompt-workflow-self-improvement/anti-pattern=Exhaustive Diminishing returns データ workflow-review 調整
Cadence worker hourly Workspace 目的の頻度 setup or workflow-review
Cadence workflow-review weekly + ad-hoc trigger Adhoc trigger 発火頻度 workflow-review 調整
Cadence digest daily User UX workflow-review 調整
Focus theme 期間 3 ヶ月 (workspace override 可) Theme apply 頻度が過剰/不足 workspace override or workflow-review
Focus theme apply gate (hard rule: Layer 3 blocking critic 必須) — 変更禁止
Layer 1 Critic 質問セット 5 問/role (rubber-duck-review.md 参照) Critic verdict skew role 単位で追加/削除、workflow-review 調整
Layer 3 Rubric 段階名 (info/minor/major/critical/blocking) (hard rule) — 変更禁止 (verdict 比較保護)
Layer 3 Rubric 各段階判定基準 現行定義 (rubber-duck-review.md 参照) Severity 分布 skew workflow-review 調整
Anti-pattern K 閾値 3 反復 誤 skip 発生 workflow-review 調整
Anti-pattern retention 90 日 Registry 肥大化 workflow-review 調整
Adapter 選択 環境依存 (Copilot Scheduler / Scout / OpenClaw / Copilot App / GH Actions / Task Scheduler / cron) Adapter 障害 / 機能変化 環境変化時に user notify + workflow-review propose
Push cadence manual (setup で 1 度質問、既定 manual) Remote 運用実態 setup or workflow-review
Throughput attempt 目標 / lane cap workspace ごと Gate 通過率の推移 workflow-review 調整 (workflow.md ## Throughput KPI)
Publish 見込み / Gate 通過率目標 workspace ごと (publish は soft forecast) 通過率が安定したときだけ attempt 増 workflow-review propose → user 承認
Blocker Test 4 問の中身 (hard rule) — 変更禁止 (中身は固定)
Approval bucket 構造 (auto / security-approve) (hard rule) — 変更禁止 (下記 invariant check)
Backup-First 原則 (hard rule) — 変更禁止
Critic 3 Layer の存在 + Independence 契約 (Layer 3 は別 model family 必須 / fail-closed) (hard rule) — 変更禁止
Review repair contract(substantive repair 後の独立 re-review / durable round record / repair の自己認証禁止) (hard rule) — 変更禁止
Layer 3 blocking gate 対象 5 種 (portfolio promote / focus theme apply / prompt-self-improvement commit / external publish / small-bet→本実装昇格) (hard rule) — 変更禁止 (rubber-duck-review.md SSOT)
Fallback lane Auto-Refill 契約 (hard rule) — 変更禁止 (中身の lane 順序は tunable)
North-star + focus theme の合意事実 (hard rule) — 変更禁止 (期間だけ tunable)

Tune Apply の詳細動作

変更 log は dashboard-state.tuningLog に append。schema は references/dashboard-state.md。

Autonomy Mode 別動作:

Mode Propose Apply Revert 追跡
Normal workflow-review が propose user 承認 (security-approve) user 判断
AUTO 既定 workflow-review が propose user 承認 (security-approve) reporter-learner が 3 サイクル追跡、悪化で自動 revert
FULL workflow-review が propose 自動 apply reporter-learner 3 サイクル追跡、悪化で自動 revert
ALL workflow-review が propose 自動 apply + criteria 拡張可 同上

Rule: どの mode でも hard rule 変更疑いは security-approve に escalate、user 明示承認まで proceed 不可。

Revert Contract

reporter-learner が 3 サイクル追跡:

  • tuningLog の trackedMetrics を各サイクルで測定
  • Baseline (change 前 3 サイクル平均) と比較
  • 悪化閾値 (既定: -10% 以上、これも tunable) 超過で revertVerdict = { verdict: "reverted", reason } を append
  • Revert は元の値を再 apply、tuningLog に新 entry (reason: "auto-revert")
  • 悪化なしなら revertVerdict = { verdict: "kept" }

Hard Rule 誤変更抑止 (Invariant Check)

workflow-review が weekly + ad-hoc で以下 invariant を check:

  1. references/approval-policy.md に "auto" と "security-approve" バケット両方の見出しが存在するか (grep)
  2. references/rubber-duck-review.md の "Layer 3 Blocking Critic" 節に 5 gate が全部あるか
  3. references/fallback-lane.md の Auto-Refill 契約節が存在するか
  4. references/persistence-profile.md の 3 profile (Standard / Persistent / Exhaustive) が全部定義されているか
  5. SKILL.md の "Tunable vs Hard Rules" 節が存在するか
  6. references/rubber-duck-review.md に "Repair -> Re-review Contract" と blocked-independence があり、runtime.limits.reviewRepairRounds が 3〜20、independenceBlockLimit が 1〜5 を維持しているか

Invariant 違反検出時:

  • dashboard-state.hardRuleViolationLog に append
  • workflow-review が blocker として user notify + revert 提案
  • Auto-apply mode (FULL/ALL) でも hard rule 違反は必ず user 承認要 (security-approve escalate)

既存 workspace への影響 (Backward Compatibility)

  • tuningLog / hardRuleViolationLog は新 field。既存 workspace の dashboard-state.json に field が無くても動作
  • 新 tuning event 発生時点で workflow-review が自動 append (field が無ければ [] で初期化)
  • 既存 tune 履歴は補完しない (前向き記録のみ)

Split Rule

このファイルが 300 行超えたら references/tunable-defaults-{category}.md に split:

  • -fallback (fallback lane / discovery floor / anti-pattern)
  • -persistence (3 profile / class mapping / attempt-log)
  • -cadence (worker / workflow-review / digest / burst)
  • -critic (layer / rubric / questions)

Split 時も本ファイルは index として残す (categories 一覧 + hard rules 再掲)。

See Also

  • references/runtime-modes.md: ai-autonomous preset SSOT (Autonomy Mode 別 tune 動作 table)
  • references/dashboard-state.md: tuningLog / hardRuleViolationLog schema
  • references/rubber-duck-review.md: Critic Layer / Layer 3 Blocking Critic SSOT
  • references/approval-policy.md: Approval bucket 構造 (hard rule)
  • references/fallback-lane.md: Fallback lane 順序 (tunable) / Auto-Refill 契約 (hard rule)
  • references/persistence-profile.md: 3 profile 数値 (tunable) / Replan 契約 (hard rule)

Source: SKILL.md on GitHub

2 warnings22d3 checks · Risk MEDIUM
  • Gen Agent Trust Hub22d

    This skill is an autonomous agent framework that manages complex workflows. It is rated as medium severity because it is self-modifying (the AI edits its own instructions), establishes persistent automated loops, and automatically accesses local secrets for its operations. While it includes many internal safety rules, the high degree of autonomy requires review before deployment.

  • Socket22d

    No alerts

  • Snyk22d

    Risk: MEDIUM · 1 issue

Signed by skilld at e63d714. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 18 hours ago.

Activeupdated 4 weeks ago
argument-hint
対象ドメイン、成果物タイプ、制約、今回の入力
user-invocable
true
metadata
{
  "author": "yamapan (https://github.com/aktsmm)"
}

README badge

README badge for aktsmm/agent-skills/opportunity-factory