메뉴
Detect and reduce overconfident reasoning in analysis, plans, estimates, reviews, and recommendations. Use when conclusions may be based on weak evidence, hidden assumptions, premature certainty, or insufficient exploration of alternatives.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Query and manage project state and drive shark workflows — epics, features, tasks, bugs, change-cards, ideas. Use for ALL project state queries, task lifecycle, status transitions (status set / advance --outcome), claim/release/heartbeat leases, blocking/unblocking, notes/context/related-docs, status dashboards, analytics, creating/updating/deleting entities, and any command starting with "shark". Also the router for `/shark <verb>` and triggers the dispatch loop on `/run <key>`.
Multi-angle parallel code review. Six specialist subagents run in parallel (A: line-by-line bugs + production caller chains, B: removed behavior + SOLID, C: cross-file contracts + structural sibling check, D: reuse/complexity/idioms, E: tests design + counter-factual, F: standards crosswalk with citations), then a consolidator verifies and ranks findings into a PASS / PASS-with-triage / FAIL report with Blocker / Non-blocker / Nit triage. Triggered by /deep-review (aliases: /comprehensive-review, /pr-review). Flags: --fix (apply one-liner safe fixes), --comment (post as inline GitHub PR comments). Prompts live in references/angle-*.md and references/consolidator.md — readable on any platform even without the Workflow tool.
Authoritative source for all validation, code review, and quality assurance workflows. Provides consistent quality gates across all development phases.
Write the test first, watch it fail, write minimal code to pass (user) - when implementing any feature or bugfix, before writing implementation code
Orchestrate User Acceptance Testing for a feature — gather evidence, run Codex as the sole assessor, present a no-opinion report to the user, record results, route rejections, and triage non-blocking findings.
Quick-capture and classify a work item discovered during development, routing it to the correct entity type (task, feature, bug, tech-debt, change-card, idea, or note) under the right parent. Searches existing entities first to avoid duplication, proposes a classification, and confirms before creating. Use when you encounter work that needs to be recorded for later, the user says "/triage", or something should be tracked but its home is unclear. This is a capture-and-classify tool, NOT a create-and-elaborate tool.
SOC 직업 분류 기준
| name | overconfidence-prevention |
| description | Detect and reduce overconfident reasoning in analysis, plans, estimates, reviews, and recommendations. Use when conclusions may be based on weak evidence, hidden assumptions, premature certainty, or insufficient exploration of alternatives. |
| when_to_use | when an answer, plan, estimate, or review needs calibration against evidence strength, uncertainty, assumptions, and plausible alternatives |
| version | 1.0.0 |
| domain | reasoning-calibration |
| inputs | [{"claim_or_recommendation":"the conclusion, plan, estimate, or assertion to calibrate"},{"evidence":"observations, sources, tests, or constraints supporting the claim"},{"decision_context":"what decision depends on the claim"},{"alternatives":"competing explanations or options already considered"}] |
| outputs | [{"selected_workflow":"one of {calibrate-confidence}"},{"calibration_report":"evidence-strength assessment and confidence recommendation"},{"assumptions":"explicit assumptions with confidence and validation method"},{"missing_evidence":"information that would materially change confidence"},{"safer_rewrite":"calibrated wording or decision framing"}] |
This skill helps an agent avoid sounding more certain than the evidence supports. It does not force indecision; it makes confidence proportional to evidence, risk, and reversibility.
Use workflows/calibrate-confidence.md for any conclusion that needs a confidence check before it informs a decision.
Always separate:
