菜单
Use when crafting prompts that require factual accuracy, working with long documents, generating claims that need verification, or when outputs must be auditable and grounded in source material
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when deciding how to manage, audit, sync, enable, disable, or troubleshoot skills and agents with the skillshare CLI across Claude Code, Codex, Gemini, OpenCode, Kiro, Goose, Pi, Antigravity, or project-local .skillshare setups. Especially use for questions about global vs project mode, disabled-by-default policies, .skillignore/.agentignore, target drift, broken symlinks, duplicate skills, sync dry-runs, or whether to use skillshare versus a runtime-native tool such as Codex skill-manager.
Use when creating git commits or reviewing commit messages. Ensures messages follow the Conventional Commits specification v1.0.0, which enables automated changelogs and semantic versioning. Invoked when the user wants to commit changes, craft a commit message, or validate an existing one.
Use when user references previous work phases, tasks, or discussions from prior sessions not present in current conversation context.
Use when designing or reviewing Python classes, deciding where logic belongs, or recognizing antipatterns like helper classes, utility bags, and god objects.
Use when transitioning between work sessions and wanting to pick up where the previous session left off.
Use when: tracking work, adding todos, filing bugs, managing tasks/issues, checking what to work on, planning epics, deferring work, pre-PR checks, or any mention of bd/beads. Covers the full bd CLI.
| name | reduce-hallucinations |
| description | Use when crafting prompts that require factual accuracy, working with long documents, generating claims that need verification, or when outputs must be auditable and grounded in source material |
Techniques to minimize factually incorrect or unsupported outputs by grounding responses in source material and building verification into prompts.
When NOT to use: Creative writing, brainstorming, or tasks where imagination is desired.
Explicitly give permission to admit uncertainty. Prevents forced confabulation.
Focus on financial projections, integration risks, and regulatory hurdles.
If you're unsure about any aspect or if the report lacks necessary information,
say "I don't have enough information to confidently assess this."
For long documents, require word-for-word extraction before analysis. Grounds reasoning in actual text.
1. Extract exact quotes from the policy most relevant to GDPR and CCPA compliance.
If you can't find relevant quotes, state "No relevant quotes found."
2. Use the quotes to analyze compliance, referencing quotes by number.
Only base your analysis on the extracted quotes.
Make responses auditable. Each claim must trace to a source quote — retract unsupported claims.
After drafting, review each claim. For each claim, find a direct quote from
the documents that supports it. If you can't find a supporting quote for a
claim, remove it and mark where it was removed with empty [] brackets.
Explicitly forbid general knowledge when source fidelity matters.
Use ONLY information from the provided documents. Do not supplement
with general knowledge.
| Technique | How | When |
|---|---|---|
| Chain-of-thought verification | Ask for step-by-step reasoning before final answer | Complex multi-step analysis |
| Best-of-N verification | Run same prompt N times, flag inconsistencies across outputs | High-stakes single-answer tasks |
| Iterative refinement | Use output as input for follow-up verification prompt | Claims that compound on each other |
| Risk Level | Minimum Strategy |
|---|---|
| Low (internal summary) | Allow "I don't know" |
| Medium (customer-facing) | Direct quotes + citations |
| High (legal/compliance) | All strategies combined + external knowledge restriction |