| name | hydrate-extract |
| description | 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.
|
hydrate-extract
Input: one source location from hydrate-discover output.
When this skill is activated:
- Determine extraction mode:
- Deterministic: manifests (parse JSON/TOML), file structure (walk dirs), config files, conventional-commit prefixes
- AI semantic: README prose, ADR bodies, PR descriptions
- For each detected fact:
- Identify the target file in working memory (see
hydrate-draft mapping table)
- Pull a one-sentence statement
- Note source provenance: file path, commit hash, line range
- Output one finding per fact as a small structured record.
Output shape
findings:
- source: package.json
target: projectOverview.md (Stack section)
statement: 'Stack is React 18, TypeScript 5, Tailwind 3, vitest.'
confidence: high
- source: src/api/types.ts
target: dataContracts.md
statement: 'User and Account interfaces define the canonical API response shapes.'
confidence: high
- source: git log
target: decisionLog.md
statement: 'Migrated from Redux to Zustand on 2026-03-12 (commit a4f8e21).'
confidence: medium
Common gotchas
- Inferring conventions from a single example. A pattern is a convention only when it shows up in multiple places.
- Treating outdated comments as current truth. Cross-check what comments claim against what the code actually does.
- Pulling debugging breadcrumbs into findings. "We had a flaky test on Tuesday" is session noise, not working memory.
