felipefontoura/axi25-core/axi25-ingest

Ingest — Deep Extraction Standard

Transform raw sources into structured, interlinked wiki knowledge. This is the core wiki-llm operation: knowledge is compiled once and kept current, never re-derived per query.

Read 90-system/references/page-templates.md for templates before creating pages.

Voice: Respond in the language the user is speaking (see 90-system/references/user-profile.md).

Standard Ingest Workflow

1. Read the source (from 10-sources/, 00-capture/, or inline).

• If it is an internal reflection/study note, route to axi25-study unless the user explicitly asked for promotion.

1. Discuss key takeaways with the user (unless they say "batch" or "just do it").
2. Create source summary in 20-wiki/syntheses/source-<title>.md using DES (see below).
3. Extract atomic concepts → create or update 20-wiki/concepts/.
4. Extract entities → create or update 20-wiki/entities/.
5. Update relevant area pages → 20-wiki/areas/.
6. Update relevant Maps of Content → 20-wiki/maps/.
7. Update index.md and append to log.md.
8. Report: what was created, updated, linked.

Optional: Zeitgeist Scouting

When the source has scan_for_zeitgeist: true in its frontmatter (or the user says "scout signals", "do a zeitgeist scouting pass", "what signals are here"), run a lightweight discourse scouting pass after the DES extraction. This is for users who work with media / need to stay tuned to the current discourse ("spirit of the age").

1. Read the source (already ingested into 10-sources/ or 50-zeitgeist/discourse/).
2. Identify 1-5 candidate signals. Five signal types (canonical English labels below; write the

stored label in the user's language — localize per AGENTS.md), consistently:

• hot-take — declarative punchy phrase, citable, polemic
• contrarian — position against current consensus
• new-frame — a new angle on a known topic
• new-data — a specific number/fact/case-study that changes the argument
• tension — disagreement between relevant voices

1. For each signal, write 50-zeitgeist/discourse/<type>/<source-slug>/signals/signal-NNN-<slug>.md

using the Discourse Signal template in 90-system/references/page-templates.md.

1. Rate each signal with pauta_rating (1-5) using this default rubric (adapt to the user's domain):

``` score = 1 (base)

• 2 if it intersects one of the user's core areas (see user-profile)
• 1 if it connects with other signals already captured
• 1 if it has a teaching element (explains something new)
• 1 if it has a narrative element (punchy phrase, story)

− 1 if it requires context the user does not command − 1 if the production cost is prohibitive − 1 if it decays fast (< 3 months) clamp 1-5 ```

1. Report a ranked table to the user. Do NOT auto-create synthesis or wiki pages

— signals live in 50-zeitgeist/, not 20-wiki/.

Routing by source type:

Deep Extraction Standard (DES)

This vault is a second brain — not a collection of connected summaries. Every source-* page in 20-wiki/syntheses/ must extract craft, not narrate facts.

Mandatory application: every mature source ingest (course, book, article, paper, podcast, promoted study note) follows this standard. Early internal brainstorms and study notes default to 25-studies/, not DES.

7 Layers per Significant Idea

For each meaningful idea in the source:

1. The idea — in your own voice, clear and precise.
2. Mechanism — WHY does it work? The underlying principle, not just the fact.
3. Source example — unfolded, not just named. What happens, in what context, how it operates.
4. Applied example — mapped to the user's context/goals when there's obvious transfer.
5. Anti-pattern or variant — how it fails, how it varies.
6. When to use / when NOT to use — contextual applicability.
7. Operational hook — how the user applies this tomorrow.

Frameworks vs Tactical Plays

• Large frameworks (3-act structure, storytelling levels, sales funnels)

get full decomposition. Each component deserves 7 layers.

• Tactical plays / individual patterns / pocket rules do NOT get 7 layers each.

The category gets it. Example: a family of related patterns in ≤100 lines covering all of them — not 7 layers per pattern.

Voice (second brain ≠ landing page)

Cold, analytical, skeptical. Cut hype adjectives and self-praise. If an idea demands emphasis, describe the mechanism; don't decorate it with adjectives ("pure gold", "brutal", "state of the art", "strong signal", etc. are forbidden crutches).

Structural Anti-Shortcuts

• NEVER consolidate 3+ chapters into one section. If a chapter has a distinct idea,

it gets its own 7 layers. If it's a variation, name it explicitly ("Chapters 5-7 apply chapter 4 to X/Y/Z, no new mechanism").

• "Surgical application to your context" is a 3-7 actionable-lines table,

not a 15-item checklist. Strong hooks > generic hooks.

• Redundancy between sources is waste. If source A already decomposed 4 phases and

source B also does, only cross-reference; don't re-decompose.

• Cross-ref between sources: OK and desired. Direct attributed quote: only if

it's in the raw of THAT source. Quote from another source: per [[source-X]]: <quote>.

Execution Discipline

• Sequential solo for single nuanced sources (course, paper, conversation where

cross-source consistency matters). For batches of full books in 10-sources/books/, use parallel — see § Massive Book Ingest.

• Human checkpoint after every delivered source. Don't chain without explicit OK.
• Verify wikilinks BEFORE writing: ls 20-wiki/concepts/ | grep <name>,

ls 20-wiki/entities/ | grep <name>. Orphans are technical debt — only link what exists (or create a conscious stub first).

• Post-write literal self-check:
• wc -l <file> for size (never estimate)
• Grep for hype crutches; if several, rewrite
• Honest simulation: "if the raw were deleted now, does the synthesis teach?"

Calibrated Size Targets

• Dense books / full courses / manifestos: 400-700 lines
• Short tactical playbooks: 200-400 lines
• Academic papers (5-15 pages): 150-350 lines
• Isolated articles / single chapters: 100-200 lines

A 50-line summary for a 45-min course = red flag. An 800-line summary for an 8-page paper = inflation.

6-Month Test (Author's Mental Ruler)

After writing: "reading only this synthesis in 6 months, can I teach the content — without reopening the original?" If not, it's shallow. Rewrite.

The test is a mental ruler for the author, not a section of the document. Do NOT include a "Depth test" section at the end (author self-validating their own work = confirmation bias). Open directly with ## Context and close with ## Connected.

Massive Book Ingest (10-sources/books/)

When there are full books in an author folder (10-sources/books/<author>/*.md, complete text) and the user wants to cover everything ("massive ingest", "all of <author>", "go massive"), do NOT go sequential-solo — parallelize with subagents, one per book.

Pattern (orchestrator = main loop):

1. ls 10-sources/books/<author>/ — map the books + their source syntheses (20-wiki/syntheses/source-*).
2. One subagent per book (Agent tool, in parallel in a single turn). For large books

(>200K), 1 agent each; small playbooks can be grouped (~6 per agent).

1. Each subagent receives the DES spec and follows the anti-drift discipline below.
2. After all finish, the orchestrator centrally: indexes the new concepts (1 block in

index.md), recomputes ghosts, updates log.md. Subagents do NOT touch index/log.

Anti-drift discipline (what makes parallel safe):

• Partition by book — each agent extracts only ITS book's concepts. No overlap.
• Check-exists before creating (ls 20-wiki/concepts/<slug>.md) — if it exists, SKIP.
• Each agent owns ONE synthesis (its book's) — updates ## Concepts Extracted there.

Never edits another agent's synthesis/concept → zero concurrent-write conflict.

• No ghosts: link only what was confirmed via ls.
• Normal DES frontmatter + voice; content in the user's language, slug per the

user profile's filename_language (default: match the user's language), faithful to the source (canonical numbers/quotes, invent nothing).

Known risk: if the session hits the token limit mid-run, agents die on the final step (report/synthesis-update). Concepts already created stay on disk (idempotent) — the orchestrator just finalizes index/syntheses/ghosts, or re-runs (check-exists skips what exists).

When to NOT Use This Skill

• "what do I know about X" → axi25-query
• "note this" (quick save without extraction) → axi25-capture
• "scout signals from [source already in the vault]" → use the zeitgeist scouting phase above (no full DES re-extraction needed)