Skip to main content

multi-agent-deliberation

Use for high-stakes decisions: architecture forks, build-vs-buy, competitive response, major refactors, or when the user wants multi-agent debate. Spawns diverse perspectives (or Codex second opinion) and synthesizes a recommendation. Do NOT use for trivial choices or every small implementation detail.

Informações da origem

Repositório
AVIDS2/avid-skill
Última atividade na origem
23 de julho de 2026 às 16:24
Idioma detectado do SKILL.md
inglês
Estrelas
1
Forks
0

Opções de instalação

Por padrão, está selecionado o prompt que primeiro revisa a origem. Você pode mudar para um comando direto ou baixar uma cópia local.

Revise os arquivos de origem

Leia o SKILL.md e os arquivos complementares exibidos pelo SkillsMP antes de decidir se vai instalar.

Exibindo SKILL.md

SKILL.md
Instruções da origem · Visualização somente leitura
name
multi-agent-deliberation
description
Use for high-stakes decisions: architecture forks, build-vs-buy, competitive response, major refactors, or when the user wants multi-agent debate. Spawns diverse perspectives (or Codex second opinion) and synthesizes a recommendation. Do NOT use for trivial choices or every small implementation detail.
# Multi-Agent Deliberation When a wrong call is expensive, prefer competing lenses — then decide. **Optional**: skip for clear choices or when the user wants speed. ## When to Use - Build vs buy / build vs wrap competitor - LangGraph vs Pi vs custom orchestration (after `agent-framework-choice` still unclear) - "Should we pivot given competitor X?" - Large rewrite vs incremental - User explicitly asks for multi-agent or second opinion ## Procedure ### 1. Frame the decision Write: - Question (one sentence) - Constraints (time, skills, stack, distribution) - Success metric - Deadline / reversibility (one-way door?) ### 2. Run 2–3 lenses (parallel when possible) Pick lenses that disagree productively: | Lens | Focus | |------|--------| | **User/value** | Does this help the buyer this month? | | **Risk/security** | What fails in prod? | | **Tech debt / YAGNI** | Simplest path (ponytail spirit) | | **Competitive** | How does this position vs named alternatives? | | **Codex/rescue** | Independent second implementation or review | Implementation options: - Claude Code `Agent` tool with distinct system prompts per lens - Codex plugin `/rescue` or `/adversarial-review` when available - Sequential self-roles only if subagents unavailable (label each role clearly) ### 3. Synthesize (you, main agent) Produce: ```markdown ## Decision question ## Options ## Lens votes (table) ## Recommendation ## Dissent / residual risks ## What would change the decision ## Proposed next experiment (smallest test) ``` Do **not** average into mush. Pick a recommendation. Surface dissent. ### 4. User owns the call Present synthesis; user confirms. Record in `docs/decisions/` or dev-log when multi-session. ## Anti-Patterns - Five agents agreeing with the same prompt (fake diversity) - Deliberating UI padding color - Hiding the recommendation behind "it depends" with no default - Skipping research gates by debating vibes only — attach evidence ## Related - `product-bootstrap`, `research-before-build`, `agent-framework-choice`
Ver no GitHub