knowledge-compile

name	knowledge-compile
description	Compile source material from docs/raw/<topic>/ into a navigable wiki node at docs/research/<topic>.md. Use when source-material has accumulated and needs summarization with cross-refs to existing feature docs / RCAs / memory.
argument-hint	<topic-path-under-docs/raw>
context	fork
agent	general-purpose

Knowledge Compile

Reads docs/raw/<topic>/, derives a wiki node at docs/research/<topic>.md with summary + sources + cross-references + key claims + open questions. Implements Phase 4 of ADR-010.

Input: topic path under docs/raw/ (e.g. tolkien/middle-earth-geography). Output: docs/research/<basename>.md where <basename> is the leaf component (middle-earth-geography.md).

What this skill does (and doesn't)

Does:

• Inventories docs/raw/<topic>/ recursively.
• Greps docs/features/, docs/reviews/, docs/ai-includes/, docs/adrs/ for keyword matches to seed cross-references.
• Synthesizes a wiki node following the structure documented in docs/research/README.md.
• Validates the output by running /lint-docs --quick after writing.
• Regenerates backlinks so the new node is wired in immediately.

Doesn't:

• Modify anything in docs/raw/. Source material is read-only from this skill's perspective.
• Edit existing feature docs, ADRs, or RCAs. Cross-references go in the research node, not retroactively in target docs (backlinks handle the reverse direction).
• Fetch new material from the web. The user is responsible for populating docs/raw/ (via Obsidian Web Clipper or manual curation).

Phases

Phase 1: Inventory

Run the briefing probe:

python tools/compile_research.py "$ARGUMENTS"

This emits a markdown briefing with:

• Full file inventory of docs/raw/<topic>/
• Keyword-matched candidate cross-references from existing docs
• A skeleton outline for the target research node

Read the output. If docs/raw/<topic>/ is empty, stop and tell the user to add source material first.

Phase 2: Read the raw material

Open each file in docs/raw/<topic>/. For text files, read the full contents (these should be small — large binaries don't belong in raw/, per docs/raw/README.md). Note for each:

• Provenance (who/what is the source, when)
• Key claims it makes
• Anything that contradicts other sources in the same topic dir

Phase 3: Audit the candidate cross-references

The probe's keyword-match candidates are candidates, not facts. For each:

• Skim the candidate doc and confirm the topic is genuinely relevant.
• Drop candidates that share only superficial keyword overlap.
• Add cross-references the probe missed — check memory file references (memory lives at the auto-memory path; reference memory entries by filename in prose).

Phase 4: Draft the research node

Write docs/research/<basename>.md following the structure in docs/research/README.md:

# <Title-Case Topic>

> Source materials: `docs/raw/<topic>/` — <N> files, last compiled <YYYY-MM-DD>

## Summary

<1-2 paragraphs: what does the raw material collectively say>

## Sources

- `docs/raw/<topic>/<file>` — <one-line description>
- ...

## Cross-references

- [features/<name>.md](../features/<name>.md) — <how the topic connects>
- [reviews/rca-<name>-<date>.md](../reviews/rca-<name>-<date>.md) — <past lesson>
- Memory: `feedback_<name>` — <relevant captured guidance>

## Key claims

- **<claim>** — source: `docs/raw/<topic>/<file>`. <Brief rationale.>
- ...

## Open questions

- <Gap in the source material>
- <Contradiction between sources>
- <Follow-up research that would strengthen this node>

Rules:

• Every claim must cite which raw file it came from (line-level if relevant). Knowledge without provenance is rumor.
• Cross-references use markdown links to existing repo files. Memory entries are mentioned by filename in prose (don't link to per-user home paths — tools/lint_docs.py will flag those).
• The Open questions section is the most important. It's how the next compile pass knows what to focus on.

Phase 5: Wire it into the wiki

After writing the file:

python tools/lint_docs.py --quick
python tools/build_backlinks.py --dry-run

If lint reports dead links in the new node, fix them before applying backlinks. Then:

python tools/build_backlinks.py

Phase 6: Surface the result

Tell the user:

• The path of the new research node
• How many sources it summarized
• How many cross-references it wired
• The open-questions list (so they can add raw material to address gaps)

Do not commit. Leave that to the user.

Re-running on an existing topic

If docs/research/<topic>.md already exists, the skill overwrites it cleanly. The user can hand-edit the Open questions section between compile runs to refine focus; everything else is bot-authored and can be regenerated.

Out of scope (don't do these here)

• Don't fetch from the web. Web material goes through Obsidian Web Clipper into docs/raw/<topic>/ first.
• Don't modify existing feature docs. If you find that a research node reveals a gap in a feature doc, surface it to the user; let them update the feature doc by hand.
• Don't lint outside the new node. Phase 5's lint pass is bounded to the dead-link probe; full-tree lint is /lint-docs.

See also

• docs/raw/README.md — input layer rules
• docs/research/README.md — output layer conventions
• ADR-010
• tools/compile_research.py — the inventory probe this skill drives