document-mind-map

Conjure the shape of a document so Nile can find the middle he keeps losing. Stolen from a claude.ai mind-map skill; the charm of the original is that it costs one step, so you stop rationing it — keep that. Default output is an in-chat nested outline you can throw away; persistence is opt-in. (Provenance + the full 10×10× review: vault KB/Claude/document-mind-map skill — Source Article + RMA-DMA Review (2026-07-01).md.)

The 3 non-negotiables (this IS the skill)

1. One core idea at the center, everything ladders up to it. (Exception: a genuinely multi-topic doc — survey, anthology — has no single thesis; map it as a forest with a scope label and N sibling centers. Never fabricate a thesis a survey doesn't assert.)
2. Flag the non-obvious cross-branch links — connections the headings don't reveal. This is where insight hides.
3. End with the questions the document never answers. This is the payload — it turns the map from a table of contents into a starting point for thinking. Never drop it. (QC below.)

Default flow (in-chat, lean)

1. Read the whole thing first — you can't find the center from the first pages. (Claude's context here holds even a 60-page PDF directly; see Large docs only if it genuinely won't fit.)
2. Find the core idea (or forest). Branch supporting points 2–4 levels deep, short labels (~6 words is a guide, not a cap — clarity wins).
3. Preserve hedges. If a claim is conditional/probabilistic/contested/single-sourced, keep its qualifier (may / under X / n=1) — never strip it to hit the word budget. A longer honest node beats a short false one.
4. Mark what you inferred. Extracted nodes are plain; anything you synthesized — the center of an ambiguous doc, a cross-link, the questions — gets an [inferred] marker so Nile can tell the document from your reading.
5. Cross-branch links (§ non-negotiable #2): each should name an anchor in both branches it joins; if it's your inference rather than stated in the text, mark it [inferred]. Don't invent connections that merely sound insightful.
6. Emit as a nested markdown outline (readable raw, folds in Obsidian). End with the questions + the one-line caveat.
7. If the input is genuinely unstructured (rambling), say so and don't force a clean tree onto mush — that honesty is a feature.

Doc-type steer (one line, don't ceremony it): center on what's load-bearing for the type — a paper's thesis (methods/results ladder up, no equal corner), a meeting's decisions/owners/unresolved (deadlines on nodes that have them), a manual's goal.

The payload — quality control (the crown jewel)

The questions must be specific, true, and earned — not filler:

• Swap test: if a question would fit under a mind map of a different document, it's filler — cut it. Name a specific claim/number/actor/mechanism from this doc.
• Answer-check: before emitting, confirm the document doesn't actually answer it (scan the source, not a summary). A question the doc answers is an error that poisons the map's credibility.
• Earn it from a gap the map exposed: a thin/unsupported branch, a hedge you preserved, a cross-link the doc implies but never closes. Note the anchor, e.g. "(from thin branch B)".
• Flex the count 2–5 by how much the doc genuinely leaves open — never pad to a quota; if only one survives, say so.

Example (paper claiming remote work raises productivity):

• ❌ "What are the long-term effects of remote work?" — generic, fits any doc → filler.
• ✅ "The study measures output over 8 weeks (§4) but calls the effect 'durable' — is 8 weeks evidence of durability, or does the thin longitudinal branch hide the real question?" — names the doc's number, ties to a thin branch.

Source pointers (where they help — not a tax)

Attach a pointer where the source has a natural anchor (p.12 / §3.2 / a quoted anchor); skip it where forcing one just adds noise (a rambling essay with no structure). Rules that keep pointers from lying:

• Pointers come from the extraction, never a model's guess. pdftotext marks pages with form-feeds, not "Page N" — so page numbers must be carried mechanically (per-page extraction with an injected [[PAGE n]] tag), never counted by a model over a long doc (it drifts, worst in the middle you care about).
• A quoted verbatim anchor (¶ "…") is the one self-verifying pointer — the default for page-less inputs (pasted docs, markdown, URLs). Use §/heading-path/line-number for structured-but-page-less docs. Never put p.N on an input that has no pages.
• Synthesized nodes are exempt — the core idea and cross-links span the doc; give them (synthesis) or a range (pp.11-13), never a fake single-page cite.
• If you can't locate a node's anchor in the text, downgrade to (approx) — a wrong pointer is worse than a missing one.

Large docs

Default and only path: read it directly — Claude's context here handles dozens of pages. ⛔ The former Qwen structure-extraction pre-pass for oversized sources is REPEALED (2026-07-31, with all of BICAMERAL); if a source genuinely won't fit, chunk it and read the chunks yourself. The old text already warned "never from Qwen's summary — that's where hedges and answers vanish"; the repeal generalised that warning after Qwen returned a confident none on 3 of 5 dense documents. For a PDF, extract text first (the pdf skill / pdftotext) preserving page boundaries — do not install an extractor unprompted; if it's missing, print the one install command and stop. For a private/authenticated URL, fetch locally — don't route private content through a cloud fetch tool.

Optional Mermaid garnish

If Nile wants the GitHub-native picture, add a Mermaid mindmap block of the top ~2 levels only (it gets unreadable deeper). Caveats: Mermaid mindmap is whitespace-driven and treats ()[]{} as shape syntax — wrap any label with special chars (incl. source pointers) in "double quotes" rather than stripping. Whether it renders natively depends on the installed Obsidian's Mermaid version — unverified on this install; verify once and fall back to the outline if it parse-errors. (Markmap is not installed here, so the outline renders as plain foldable markdown, not an interactive Markmap — don't claim otherwise.)

Opt-in SAVE (only when Nile asks "save this")

Route the write through scripts/emit_map.py — it owns the mount-liveness check (non-empty vault dir), deterministic filename (YYYY-MM-DD--<slug>--<source-hash>.md, stable source-date), full house frontmatter, idempotency (skip/refresh if it already exists), and atomic write + read-back. Determinism lives in that script, not here.

• Route by content sensitivity: default to your general/technical notes. ⭐ A map of a sensitive document is exactly as sensitive as the document — the compression does not launder it. So a map of private material goes wherever that material already lives, confirm before saving, and never index it into a shared or cloud-backed memory layer.
• Wire it into the graph: on save, add related: [[…]] wikilinks + a ## Related notes block pointing at vault notes the branches touch (reuse the cross-link detection, aimed outward) — so the map isn't an island.
• Route the payload (offer, don't auto-fire): offer each open question as a Todoist capture; flag a cross-branch link that recurs across maps as an atom candidate (Neo4j via atom-cypher-safe). Your Shadow Offload Rule — don't let the cleverest output die in a file.

Report back

Where it landed (or "shown in chat, not saved"), the core idea in one line, the questions, and any spot too unstructured to map cleanly.

Deferred (not built — add only if the happy path proves too small)

Branch-expand into linked child sub-maps · multi-doc / comparison "diff map" · a Mind Maps/_index.md MOC · source-hash staleness detection · a shallow/deep depth knob. All flagged by the completeness pass; none earns its complexity yet.