| 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.
|
hydrate-draft
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.
Output
One drafted change per target file, grouped by file.
Common gotchas
- Bloating
projectOverview.md. The file is for orientation, not exhaustive cataloging. Keep it under a screen of content.
- Putting recent decisions in
conventions.md. Decisions go in decisionLog.md; conventions are stable patterns.
- Generating
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.