跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

goat-critique

Use when a decision or analysis needs multi-lens critique to surface blind spots before shipping.

blundergoat · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add blundergoat/goat-flow --skill goat-critique

技能说明


name: goat-critique description: "Use when a decision or analysis needs multi-lens critique to surface blind spots before shipping." goat-flow-skill-version: "1.16.0"

/goat-critique

Shared Conventions

Read .goat-flow/skill-docs/skill-preamble.md and .goat-flow/skill-docs/skill-conventions.md for shared conventions before proceeding.

When to Use

Use when a concrete artifact deserves multi-perspective critique before shipping: plan, security assessment, debug hypotheses, review findings, test strategy, architecture proposal, or refactor approach.

Boundary Commands

  • NEVER: Replace delegated critique with inline role-play, skip required phases, or auto-apply recommendations.
  • ALWAYS: Treat explicit invocation as consent to run the full delegated protocol against a concrete artifact.
  • DEFER TO: Before invocation only, create a missing artifact first, answer simple facts directly, or use /goat-review for a trivial artifact; explicit /goat-critique still runs in full.
ExcuseReality
"The artifact is trivial - a quick critique would cover it"Quick mode was tried and removed. A single reviewer running lens passes in one context is self-talk under three labels, not multi-perspective critique.
"All three agents agree so it must be right"Consensus without orchestrator verification is unverified self-declaration. The orchestrator's job is to verify claims, not count votes.
"Inline role-play is faster than spawning agents"Agents that role-play SKEPTIC/ANALYST/STRATEGIST inline produce indistinguishable perspectives. Isolated context is what makes findings independent.
"Closing checks happen after the main answer - skip them"End-of-task rules have near-zero voluntary compliance. Phase 5.5 meta-audit and outcome capture exist because post-deliverable steps get skipped.

Direct invocation is binding. $goat-critique or /goat-critique runs Phases 1-5 plus mandatory post-synthesis steps (5.5, 5.6). Dispatcher ambiguity rules do not override direct invocation; raise scope concerns after synthesis.

Report-only by default. $goat-critique make X shorter = critique only; $goat-critique ... then apply it = critique first, apply after gate. See Constraints for mutation and apply rules.

Step 0 - Intake

goat-critique runs only full delegated mode: Phases 1-5, 5.5 meta-audit, 5.6 outcome capture, three critique sub-agents, one meta-agent. Lighter-mode suggestions are the failure this design prevents.

Intake checklist:

  • Confirm the artifact exists and is concrete (a file, a plan document, a specific set of findings - not a vague idea).
  • Select the critique rubric for the artifact type (see Critique Rubrics below). If unclear, ask the user.
  • Use the preamble's learning-loop retrieval on relevant .goat-flow/learning-loop/footguns/ and .goat-flow/learning-loop/lessons/; record explicit misses instead of broad-loading buckets.
  • Delegation consent: proceed directly to Phase 1. Skill-chained entry: skip intake confirmation, use caller context; still run retrieval + rubric selection. All phases (1-5 + 5.5 + 5.6) always run.
  • Differential mode detection: If .goat-flow/logs/critiques/ has a same-artifact slug within 30 days, offer differential mode: A/B receive prior log + artifact diff; C stays cold. Phase 5 adds delta counts and [diff-of: <prior-uuid>].
  • Read context map: Merge the selected rubric map from references/rubric-examples.md into the fixed A/B/C split; never replace baseline context.

Phase 1 - Generate Competing Critiques

Spawn all three sub-agents in parallel using the host's real delegation mechanism.

Context varies intentionally - informational diversity catches more than tonal diversity.

The Core Trio Lens

Agents A and B both use the SKEPTIC/ANALYST/STRATEGIST combined lens. These three perspectives work as a unit - never split them into separate agents:

  • SKEPTIC - "What could go wrong? What assumptions are unproven? What's the worst-case scenario?"
  • ANALYST - "What does the evidence actually say? What's the cost/benefit? What do the numbers and code paths tell us?"
  • STRATEGIST - "What's the fastest path to shipping? What can we defer? What's the highest-leverage change?"

A/B must use all three perspectives. Agent C is the control group: same fields, no per-lens quota; use N/A - fresh-eyes scope only when a lens adds nothing.

Context split:

AgentReadsDoes NOT read
A (Risk)artifact + architecture.md + targeted INDEX-first footgun/lesson hits + rubricgit history, config.yaml
B (Alternatives)artifact + architecture.md + git log --oneline -20 + config.yaml + rubricfootguns, lessons
C (Fresh Eyes)artifact + rubric ONLYeverything else (isolation enforced)

Sub-Agent Definitions

Full directives: references/sub-agent-directives.md.

  • A (Risk): SKEPTIC/ANALYST/STRATEGIST on risks, 2nd-order impacts, fastest safe path. Must cite downstream files by name.
  • B (Alternatives): SKEPTIC/ANALYST/STRATEGIST on alternatives, ranked by implementation friction. Must surface at least one alternative.
  • C (Fresh Eyes): No project context. Flags unstated assumptions and readability gaps. ISOLATION RULE enforced.

Each sub-agent normally returns 3-7 evidence-backed findings plus required lens fields, severity, evidence, confidence, Proof class, rubric dimensions, overall assessment, and one strength. Zero findings are valid only through the reference pack's clean-result attestation after one documented second pass.

Lens coverage: A/B must analyse every lens and C must probe assumptions/readability. A lens with no supported issue gets one documented second pass, then records No supported finding instead of inventing one. See the reference pack.

Phase 2 - Rank and Compare

Execute in this order:

1. Context leak scan. Grep Agent C output for .goat-flow/, `

阅读完整来源 ↗