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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/caozx1110/ResearchLab --skill idea-workbench命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
统一分析 paper、repo、dataset、blog 与技术笔记 unit;按 kind 路由到既有 evidence-first prepare/Agent-fill/verify 实现,保持历史 owner、偏好与确认回执身份不变。
kb 快捷命令入口(伪 CLI),用于把常用 research 操作统一成 kb 动词形式;当用户运行或对 AI 说 kb help/init/doctor/update/obsidian/status/next/find/add/ingest/review/reject/recall/resume/undo/restore 时使用。
管理 core 研究系统配置,包括资源画像、语言偏好、taxonomy seed、candidate pool policy 与自动化开关。
基于 SOC 职业分类
正在显示 SKILL.md
| name | idea-workbench |
| description | 负责 core idea unit 的生成、evidence-first analysis、陪练讨论、多候选管理、显式选择与归档。 |
协议参考:
.agents/lib/research/SCHEMAS.md#unit-record·#config-files·#confirmation-gate·#runtime
当任务是把研究方向收敛成可评审、可比较、可显式选择的 idea unit 时,使用这个 skill。
select-best 或 select;selection 仍保持 pending_user_confirmation,不会被脚本自签为 confirmed。evidence_ref 格式;描述性计数明确不是 score / verdict。validate_claims,再按 source_unit_id 定位 canonical KB unit,并用 verify_claim_evidence 对该 unit 内 artifact 做逐字 quote 校验。pending_user_confirmation。证据校验通过只代表 grounded,不等于用户确认判断。generate / analyze / review / discuss 是真实 task-scoped preference consumer。prepare 暴露 value-free canonical task context;runtime Agent 可选择相关 soft preference,verify 重算 current record/request、immutable orientation 与冻结 evidence corpus 后再接受 receipt。无 receipt 时不读取 soft profile。selection_id / selection_digest / task_context_digest / skill / operation;不复制 preference value。偏好不能覆盖当前用户明确给出的题目、scope、资源边界,也不能削弱 evidence/confirmation。generate 采用 prepare|verify 两阶段合同:
prepare 只保存用户原始 title/problem/hypothesis/source/pool context、不可变 orientation、冻结的 canonical KB corpus binding 与指定数量的空候选槽位。verify 先重算 request/orientation/corpus 与可选 preference receipt,再完整验证所有槽位、唯一 identity、边界和 distinctness;全部通过后在一个 transaction 内创建所有 idea records 与 bundle。任一候选失败都零 candidate/bundle 写入。analyze 与 review 都采用 prepare|verify 两阶段合同:
prepare 生成四条空白 judgement claims:novelty、feasibility、recommendation、killer-question。evidence_refs。verify 拒绝空证据、找不到的 source unit、不可读 artifact、伪造 quote 或错误 PDF page locator;全部通过才持久化。review 可由 agent 填正整数 selection_rank,供 select-best 消费。新 review 不生成 heuristic score_breakdown;旧记录中已持久化的 score 仅作兼容读取。link --from-id <idea-id> --to-id <unit-id> --relation evidence-for 建 canonical link,再在同一 analyze/review prepare 加 --refresh-corpus;它保留已填白名单字段并重建证据边界。默认待填文件:analyze=analyze-fill.yaml,review=review-fill.yaml,discuss=discussion-fill.yaml;--input 只传 unit 根下 basename。
scaffold 中 idea_context 即使为空也由 owner 管理、只读;Agent 只填写 claim/reviewer/rank 等明确列入白名单的判断字段与 evidence。
discuss(别名 spar)采用 prepare|verify|confirm|reject 合同,并按 conclusion 粒度持久化:
prepare 生成一份空白 conclusion,包含 challenge、probe、counter-example、constructive-suggestion 与 conclusion 五条 judgement claims。verify 对每个 evidence ref 到其 source_unit_id 的 canonical unit 中核验;例如 counter-example 引用 paper 时,quote 必须逐字存在于该 paper unit 的 artifact。discussion-judgements.yaml 追加独立 idea_discussion_conclusion subject,包含 canonical payload.claims + payload.verification;nested conclusion 只是人类可读 projection。confirm 只确认指定 conclusion subject 的当前 receipt;reject 只关闭同一个已核验 subject,不要求确认署名,也不覆盖其它轮次。多轮 spar 互不覆盖既有讨论历史或已确认 analysis/review claims。payload.discussion.conclusions[] schemapayload:
discussion:
conclusions:
- id: discussion-<digest>
conclusion: agent-authored synthesis
reviewer: runtime-agent-or-human-id
verified_at: ISO-8601 timestamp
verification: evidence_verified
judgement_id: discussion-<digest>
confirmation_status: pending_user_confirmation|confirmed|rejected
claims:
- id: challenge|probe|counter-example|constructive-suggestion|conclusion
role: challenge|probe|counter-example|constructive-suggestion|conclusion
text: agent-authored judgement
claim_type: inference|evaluation
confirmation_status: pending_user_confirmation
evidence_refs:
- source_unit_id: canonical KB unit id
artifact: unit-relative artifact path
locator: page=N|section|anchor|file:line
quote: short verbatim
verified_at / verification 只代表证据核验;claim 的 epistemic type 与 pending_user_confirmation 保持不变,直到显式用户确认。
以下命令只供 agent 内部执行,不直接展示给 end user;面向用户只输出自然语言或 kb <verb>。
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py capture --title "retrieval-aware code assistant"
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py generate --title "adaptive retrieval policy" --count 4 --pool current-ideas --phase prepare
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py generate --title "adaptive retrieval policy" --count 4 --pool current-ideas --phase verify
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py analyze --idea-id i-example-f7e91d86 --phase prepare
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py analyze --idea-id i-example-f7e91d86 --phase verify
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py review --idea-id i-example-f7e91d86 --phase prepare
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py review --idea-id i-example-f7e91d86 --phase verify
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py discuss --id i-example-f7e91d86 --phase prepare
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py discuss --id i-example-f7e91d86 --phase verify
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py review-assist --pool current-ideas
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py select --idea-id i-example-f7e91d86 --confirmed-by research-lead --evidence kb/programs/example-program/decision-log.md
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py select-best --pool current-ideas --confirmed-by research-lead --evidence kb/programs/example-program/decision-log.md
${RESEARCH_PYTHON:-python3} .agents/skills/idea-workbench/scripts/idea.py archive --idea-id i-example-f7e91d86