← ClaudeAtlas

decisionlisted

Settle one contested choice on the record: name the alternatives that are really live, put the evidence for each next to it, and land a call with the reasoning attached. Use it when three or more options are viable, when the evidence points both ways, or when the choice is expensive to undo — "Postgres or DynamoDB for this?", "should we roll our own or use library X?", "is it worth migrating off Y?", "the team can't agree on A versus B", "we tried this before and it didn't work, what now?". The one most often missed is "what do you think we should use for the queue?": phrased as an opinion, so the agent just answers, and months later nobody can reconstruct why. There is a floor — a cheaply reversible choice with an obvious answer does not need this. Do NOT use it for a document's worth of connected choices (`/design`), for options nobody has named yet (`/explore`), or for surveying a field of vendors before choosing among them (`/comp`).
tsukumogami/shirabe · ★ 0 · AI & Automation · score 67
Install: claude install-skill tsukumogami/shirabe
!`bash ${CLAUDE_PLUGIN_ROOT}/scripts/skill-preflight.sh decision 2>&1 || true` @.claude/shirabe-extensions/decision.md @.claude/shirabe-extensions/decision.local.md # Decision Skill Make well-reasoned, auditable decisions through structured evaluation. The skill produces decision reports that map directly to design doc Considered Options sections and standalone Decision Records (ADRs). **Writing style:** Read `skills/writing-style/SKILL.md` for guidance. ## Decision Tiers This skill handles Tier 3 (standard) and Tier 4 (critical) decisions. For Tier 1-2, use the lightweight decision protocol (`references/decision-protocol.md`). | Tier | Path | Phases | When | |------|------|--------|------| | 3 (standard) | Fast | 0, 1, 2, 6 | 3+ options, needs research, but not adversarial | | 4 (critical) | Full | 0, 1, 2, 3, 4, 5, 6 | Irreversible, high-stakes, contested | ## Agent Hierarchy When invoked as a sub-operation by a parent skill (e.g., /design), the decision skill runs as a **decider agent**. The decider spawns sub-agents: ``` Level 1: Parent skill (design, explore) └── Level 2: Decider agent (this skill, one per decision question) ├── Level 3: Research agent (Phase 1, disposable) ├── Level 3: Alternative agents (Phase 2, disposable) └── Level 3: Validator agents (Phases 3-5, persistent via SendMessage) ├── Phase 3: argue FOR their alternative (bakeoff) ├── Phase 4: receive peer findings, revise position