brain-consolidatelisted
Install: claude install-skill layersistem/brain-kit
# brain-consolidate
Decision records pile up. Reusable knowledge stays buried inside them, older records that a
newer decision has invalidated are never marked, and contradictions sit there unnoticed. This
skill turns that cleanup into a reviewable proposal.
## Flow (3 steps)
1. **Build the prompt**
```bash
python3 "$BRAIN_ROOT/scripts/brain_consolidate.py" --since YYYY-MM-DD --until YYYY-MM-DD
```
Writes `<vault>/_drafts/consolidation_<until>_prompt.md`. Default window: yesterday..today.
No arguments needed for the normal nightly run.
2. **Write the report.** Hand the prompt file to a subagent - use the strongest model you have
available, since this is a judgement task, not a formatting task. The subagent reads the
prompt file and writes the five sections (DISTILL, SUPERSEDE, CONFLICT/DUPLICATE, WEIGHT, SKILL DISTILLATION
CANDIDATES) to `<vault>/_drafts/consolidation_<until>.md`. It touches no other file.
Gate the result before showing it:
```bash
python3 "$BRAIN_ROOT/scripts/brain_consolidate.py" --verify <report.md>
```
That checks all five sections exist and that no `[[wikilink]]` points at a missing note.
3. **Show the human the report.** Only approved items get applied:
- distil -> write or edit the target note in `knowledge/` or `memory/`
- supersede -> add `status: superseded` and `superseded_by:` to the old note
- weight -> add `weight:` to the decision record's frontmatter
Nothing is applied without appro