菜单
Map findings into the six working memory files: projectOverview, decisionLog, dataContracts, conventions, openQuestions. Drafts include source provenance inline. activeContext.md is excluded; the developer drives that file directly.
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
Inventory the source surface for hydrating working memory in this repo. Surveys manifests, code structure, README, ADRs (if any), recent git history, and code patterns. Run once when populating working memory beyond the scaffold prompt's pre-population.
Pull findings from a single source location. Each finding is one sentence about a fact, decision, convention, or contract, plus a pointer back to source. Deterministic for structured sources, AI semantic for prose.
Stage reconciled drafts as a single commit (or PR for multi-developer projects) that updates the working memory files in a reviewable batch. activeContext.md is excluded. After acceptance, the synchronizer agent takes over for ongoing maintenance.
Compare drafts against existing working memory and against the codebase. Annotates each draft as net-new, would-overwrite, or conflicts-with-code. Greenfield projects skip the existing-working-memory comparison.
Reads the current working memory state, diffs it against recent git changes, and proposes updates. Use when finishing a feature, resolving a decision, or when active context feels stale.
| name | hydrate-draft |
| description | Map findings into the six working memory files: projectOverview, decisionLog, dataContracts, conventions, openQuestions. Drafts include source provenance inline. activeContext.md is excluded; the developer drives that file directly. |
Input: findings from hydrate-extract.
When this skill is activated:
For each finding, route to the right working memory file:
| Finding type | Lands in |
|---|---|
| Stack, framework, language, deployment | projectOverview.md (Stack section) |
| Repository structure, monorepo rules, off-limits areas | projectOverview.md (Repository Structure, Key Constraints) |
| Decisions made (with context) | decisionLog.md |
| Recurring patterns: naming, file organization, error handling | conventions.md |
| Type definitions, API shapes, schemas | dataContracts.md |
| Unresolved questions about the project's intent | openQuestions.md |
Compose draft content for each target section. Keep entries terse; working memory rewards brevity.
Include source as inline reference. For decisionLog.md, cite the commit hash or ADR file. For conventions.md, cite an example file path. For dataContracts.md, point at the source type definition.
Do not write to the working memory files directly; output drafts as proposed changes for hydrate-reconcile.
One drafted change per target file, grouped by file.
projectOverview.md. The file is for orientation, not exhaustive cataloging. Keep it under a screen of content.conventions.md. Decisions go in decisionLog.md; conventions are stable patterns.openQuestions.md entries for things the AI just doesn't know. Open questions should be things the project's authors haven't resolved, not things the AI can't determine. Phrase them in the project's voice, not the agent's.