synthesise

name	synthesise
description	Complete the interactive half of the two-tier capture pipeline. Finds all status:raw source files, cross-links them to topic indexes, checks synthesis files for updates, and marks them status:summarized. Owner-triggered. Run with full KB context.

Synthesise

The Worker wrote raw source files (summary, quotes, takeaways already present). This skill does the KB integration work the Worker explicitly skipped: cross-linking, index updates, synthesis checks.

Trigger

• /synthesise — process all status: raw captures
• /synthesise <topic> — process only captures tagged with that topic (e.g. /synthesise prompting)

Steps

1. Find raw captures

Grep sources/*.md (not sources/skipped/) for status: raw in frontmatter. List them with signal level and topics before doing anything else.

If no raw captures found: say so and stop.

2. Confirm with owner

Show the count and file list. If >5 files, flag the batch size and ask whether to proceed in full or filter by topic first. Wait for confirmation before processing.

For large backlogs (>~15 raw files), offer fixed batches of N (default 10) processed in priority order (high-signal + newest first). State of done-ness lives in the source frontmatter (status: summarized), so batches are resumable: after a batch, stop and report N/M done; a fresh /synthesise re-greps status: raw and continues. This lets the owner /handoff to a clean context between batches. Within a large batch, do not rewrite synthesis bodies inline (step 3c) — collect synthesis-update candidates and surface them for confirmation at the end of the batch, alongside the open-thread scan.

Use node worker/scripts/append-source-link.mjs <topic> <source-basename> "<hook>" to append a ## Sources bullet to a topic index (LF-safe, idempotent — skips if the source is already linked). This avoids re-reading large indexes for every cross-link; you still author the one-line hook.

3. Process each source (one at a time — finish each before starting the next)

a. Read the full source file. Do not work from memory.

b. Cross-link to topic indexes. For each topic in the source's topics: frontmatter:

• Read topics/<topic>/index.md. If it doesn't exist, create it with a minimal header.
• Add a one-line entry under ## Sources (or the most specific matching section if the index has one, e.g. ## Synthesis for synthesis pointers). Format: - [Title](../../sources/YYYY-MM-DD-slug.md) — one-line hook
• For low-signal sources, append [low-signal] to the entry and do not add to open threads.

c. Check synthesis files. Scan synthesis/ for any file whose topics: frontmatter overlaps with this source. If found:

• Read it and assess whether the new source changes it — new term, new evidence, new failure mode, new framing.
• If yes: update the synthesis body, add the source to its sources: frontmatter list, update updated: to today.
• If no: leave it untouched.

d. Mark as processed. Edit the source file's frontmatter: status: raw → status: summarized.

4. After all sources — cross-link new indexes

If this run created any new topic indexes, scan them for conceptual overlap with each other. For any pair that clearly covers adjacent or overlapping ground, add a ## See also section to each with a one-line pointer to the other.

Example pairs that warrant cross-links: agent-orchestration ↔ agent-architecture, agentic-workflows ↔ workflow-automation, memory-systems ↔ memory.

Only add cross-links where the overlap is concrete and navigable — not every adjacent topic. Skip if the batch created fewer than 3 new indexes.

5. After all sources — open thread scan

Check the Open threads sections in every topic index touched during this run. If any thread now has ≥3 supporting sources or shows a clearly resolvable disagreement across ≥2 sources, surface it briefly:

"Open thread in topics/<topic>/index.md: [thread name] now has N sources — worth writing the synthesis?"

Do not write synthesis unprompted. Ask first.

Constraints

• Read every file — never recall content from memory.
• Do not re-summarize content already in the source file. The Worker wrote the summary, quotes, and takeaways.
• Do not write synthesis unprompted. Flag and ask.
• Low-signal sources still get indexed but with a [low-signal] marker; do not use them to open new threads or promote new tags.
• Finish one source completely (steps a–d) before moving to the next.
• If context budget is getting tight mid-batch, stop after the current source and report progress: "processed N/M, stopped to avoid context overflow — run /synthesise again to continue."