LLM Knowledge Base

Skill: compile-ingest

Follow the compile checklist in AGENTS.md. This skill fills in the what behind each step.

fetch-url: command

When you receive fetch-url: <url> [name]:

./tools/fetch-url.sh "<url>" [name]

The script auto-detects static vs JS page:

• Static page (≥200 words) → uses requests + markdownify — fast, token-efficient
• JS-rendered page (<200 words) → falls back to r.jina.ai automatically
• Force JS: ./tools/fetch-url.sh "<url>" --force-jina

After the script saves raw/articles/<name>.md, continue with the normal compile checklist from Step 3.

Step 3 — Image Handling (before reading)

Check for external images first: grep -c 'https://' "raw/articles/file.md"

If count > 0 → run ./tools/fetch-images.sh "raw/articles/file.md" before reading. Without this, external images show as broken URLs — Claude cannot see them and will write wiki content with missing visual context.

Image triage when writing wiki:

• HIGH VALUE (diagrams, flowcharts, architecture, data charts) → describe in detail + embed ![name](../../raw/images/name.png)
• LOW VALUE (banners, avatars, stock photos) → ignore completely, do not mention

Step 4 — Core Cognitive

Summary format (wiki/summaries/<name>.md)

Required sections in this order:

1. Executive Summary — 1–2 paragraphs: context, problem, main thesis
2. Deep Analysis — logical breakdown with sub-headings
3. Key Insights — 2–5 bullets of the most novel/counter-intuitive findings
4. (optional) Limitations & Open Questions
5. Concepts + Relations — wikilinks at the end

Write to be understood, not just to record. Explain why something matters, not just what it is.

Concept extraction (wiki/concepts/<name>.md)

Hard limit: 1–3 MACRO concepts per document. Minor terms and sub-concepts must be grouped under a parent macro — do not give them their own files.

Required frontmatter: domain: field is non-negotiable. Use the domain MOC slug (ai, product, meta...). If no MOC exists yet: use a short slug, lint will detect it.

Compiled Truth + Timeline format — every concept file ends with:

---
<!-- TIMELINE (append-only — do not edit existing entries) -->
- YYYY-MM-DD: Source — one line on how/why compiled or updated

• Above --- = compiled truth (rewrite when evidence changes)
• Below --- = timeline (append-only, never edit existing entries, only add new)

Long document strategy

Run ./tools/scan.sh --info <file> → check word count + section map → pick strategy:

MAP-REDUCE — correct flow (do NOT chunk by line count):

1. ./tools/scan.sh --info <file> — output now includes section map with exact line boundaries
2. Chunk by section boundaries from the map, skip References/Appendix/Proofs
3. Spawn agents in parallel, one per logical section group
4. Each agent: read its lines → produce structured notes (NOT a full summary)
5. REDUCE: combine all notes → write final summary + concepts

Cost trade-off: MAP-REDUCE costs ~3–4x more tokens than sequential. Use only when context overflow is a real risk (> ~15K words in dense technical text). For most papers and articles, sequential read is faster and cheaper.

Step 4.5 — Verify before finalizing

• Summary has all required sections?
• Concepts ≤ 3 macro, each with domain:?
• All [[wikilinks]] point to files that exist (or will exist after this compile)?
• No unexplained jargon stacking?

If any check fails → fix first, then proceed to Step 5.

Step 5 — finalize-compile.sh

./tools/finalize-compile.sh "raw/path/to/file" "One-line key insight" --model claude-sonnet-4-6

--model is required — pass your actual model ID. Supported IDs: see tools/metrics.py.

If no domain MOC exists yet and ≥10 concepts exist → manually create MOC following wiki/domains/_about-domains.md.

Gotchas

Concept explosion — The most common mistake. After reading an interesting document, the instinct is to create a concept for every interesting term. Resist. Ask: "Is this a MACRO idea that stands on its own, or a sub-concept that belongs under an existing concept?" When in doubt, group it.

Missing --model flag — finalize-compile.sh runs silently without it but cost tracking is broken. The flag is not optional; it's required for metrics.

Editing TIMELINE entries — The <!-- TIMELINE --> section is append-only. Never edit existing dated entries, even to fix typos. Only add new dated lines at the bottom.

Skipping fetch-images.sh — If you read a raw .md file before downloading its images, external URLs appear as broken links. You'll write wiki content without seeing the actual diagrams. Always grep for https:// first.

Choosing strategy without scan.sh --info — Don't guess the word count. A "short article" might be 12,000 words. Running scan.sh --info takes 2 seconds and determines the right strategy.

Concept file > 150 lines — If you're going over, the concept is doing too much. Split into parent + sub-concept, or move examples to a summary instead.

Not verifying [[wikilinks]] before finalizing — Orphaned links will fail lint. Check that every [[linked-concept]] either exists already or will be created in this same compile session.

Source: hoadoan1997/llm-knowledge-base — distributed by TomeVault.