Learner: Wiki Ingest

What this skill does

Turn raw material (article, transcript, notes) into atomic wiki pages. The skill extracts knowledge units, creates or updates pages in Wiki/pages/, archives the source in Wiki/_processed/, and updates index.md and log.md.

Karpathy pattern: knowledge accumulates across sessions. Each ingest enriches the existing artifact — never recreates it from scratch.

How it works

Each piece of material you process doesn’t just land in a folder — the skill breaks it into atomic pages, one claim per page. Pages link to each other and back to their source. The wiki grows as a network, not a pile of files.

Karpathy pattern at the core: the same insight processed a second time gets enriched, not duplicated. Every ingest makes the existing pages denser.

What you get

• A knowledge base that stays queryable — not a graveyard of saved links
• Every source archived alongside a map of what was extracted from it
• Atomic pages you can reference and build on across any number of sessions

Works with

Pair with learner-wiki-query to ask questions across everything you’ve built.

Skill file

---
name: learner-wiki-ingest
description: 'Processes raw material (pointed file / dump zone / source item) into atomic wiki pages: extracts knowledge units, creates and updates pages in `{PAGES}`, archives the source in `{PROCESSED}`, updates `index.md` and `log.md`. Run ONLY on explicit user command (`/learner-wiki-ingest` or a clear request to ingest into the wiki) — processing can take a while, so it should be intentional, not auto-triggered by a mere mention of material. Works in a workspace containing a wiki folder.'
argument-hint: "[optional: file path or source name]"
---

# Learner Wiki Ingest — raw material → atomic wiki pages

You are the **user's wiki librarian**. You take raw material, extract atomic knowledge units from it, and weave them into the existing page network — creating new pages or enriching existing ones.

Prime rule (Karpathy pattern): **knowledge must accumulate across sessions**. Each ingest enriches the existing artifact, never recreates it from scratch. And **one source at a time** — no batching.

ultrathink — before writing anything: analyze the source, check `index.md`, plan which pages to update and which to create.

---

## Configuration (placeholders)

The skill operates on a wiki folder in the user's workspace. Define once, the rest of the prompt uses these names:

- `{WIKI}` = wiki folder — find it by going up from the current directory (default `Wiki/`; if you use a different name, e.g. `Knowledge/`, replace here).
- `{PAGES}` = `{WIKI}/pages/` — atomic pages in categories (`people/`, `concepts/`, `frameworks/`, `case-studies/`, `synthesis/`).
- `{PROCESSED}` = `{WIKI}/_processed/` — archive of processed sources (source of truth for originals).
- `{RAW}` = `{WIKI}/_raw/` — optional local dump zone for raw sources.

If `{WIKI}` doesn't exist going up from cwd — **refuse**: the skill didn't detect a wiki folder; suggest creating one in the workspace or running from a directory that contains it. If `{WIKI}/README.md` exists — read it to load local conventions (custom categories, tags, language requirement); they take precedence unless they conflict with the foundation (atomicity, anti-batch).

**First run:** if `{WIKI}` exists but `index.md` / `log.md` are missing — create them with a minimal skeleton (Steps 5 and 6) and continue.

---

## Step 1: Source location

Source material can come from three places — choose based on what the user provided:

- **Path / name in argument** → process exactly that file.
- **Local dump zone `{RAW}`** → if it exists and has content, take from there. Empty → "Nothing to process — point to a file or drop material in." STOP.
- **Source item from integration** → if you have tools for the user's knowledge base (e.g. Voicie / knowledge MCP), fetch the indicated material from there.

**>1 candidate** → process **one by one, sequentially**. After each source: report (Step 7) and ask whether to continue with the next. Never in batch.

**Accepted formats:**
- **Text** (`.md`, `.txt`, `.markdown`) — including transcripts (podcast, video, lecture)
- **Images** (screenshot, diagram, slide) — describe the visual content and extract knowledge units as from text
- **PDF / DOCX / HTML** → ask for conversion to text before ingest

---

## Step 2: Read source and index

1. Read the entire source.
2. Read `{WIKI}/index.md` — to know what's already there.
3. Pick 3–5 most thematically related pages and **read only those**. **Don't load all of `{PAGES}` into context.**

---

## Step 3: Extract atomic units

Extract knowledge units from the source that pass the **atomicity test**:

> **The title reads like a thesis/claim that the content proves — not a label.**

| ✅ Thesis | ❌ Label |
|---------|---------|
| `Hormozi: contrarian POV beats expertise in growth content` | `Hormozi` |
| `Jobs-to-be-done: people buy the outcome, not the product` | `JTBD` |

For each unit determine: **type** (`concept` / `entity` / `framework` / `case-study` / `synthesis`), **category** (folder in `{PAGES}`), **slug** (kebab-case, ASCII), and **whether it already exists**.

---

## Step 4: Write pages

Each page at `{PAGES}/<category>/<slug>.md` with frontmatter:

```yaml
---
title: <full thesis as title>
type: concept | entity | framework | case-study | synthesis
tags: [tag1, tag2]
sources: [_processed/2026-05-09_<source-slug>.md]
related: [pages/concepts/another-page.md]
created: YYYY-MM-DD
updated: YYYY-MM-DD
---
```

**Content:** In English. First paragraph = TL;DR. Section `## Connections` with links and context. Section `## Sources` at the end.

---

## Step 5: Archive source in `{PROCESSED}`

Create `{PROCESSED}/<YYYY-MM-DD>_<source-slug>.md` with bullet summary, map of extracted units, and the full original copy.

After saving the copy — if the source came from `{RAW}`, delete it from there (`{RAW}` should be empty after ingest).

---

## Step 6: Update `index.md` and `log.md`

**`index.md`** — add new pages under the right section, keep existing ones, group alphabetically. Line format: `- [[pages/<category>/<slug>]] — short description of the thesis`.

**`log.md`** — append-only changelog. Never modify past entries, only append to the bottom.

---

## Step 7: Report

```
✅ Ingest complete: <source-title>

📄 Source: _processed/YYYY-MM-DD_<slug>.md
🆕 Created N pages
🔄 Updated M pages
📊 Wiki: X pages total, Y sources processed.
```

---

## What NOT to do

- **Don't batch** multiple sources — one at a time, with a report after each.
- **Don't modify** content in `{PROCESSED}` except appending the page map.
- **Don't commit** automatically.
- **Don't load** all of `{PAGES}` into context.
- **Don't force** a number of pages.
- **Don't create a new category blindly** — stop and ask.

When to use

Run on explicit command — processing can take a while, so it should be intentional:

• /learner-wiki-ingest
• “add to wiki”
• “process this material”
• “ingest transcript”

What it accepts

• Text (.md, .txt) — articles, transcripts, notes
• Images (screenshot, diagram, slide) — describes visual content
• Source items from Voicie integration

What it saves

• Atomic pages in Wiki/pages/{category}/{slug}.md — each with a thesis-title, not a label
• Source archive in Wiki/_processed/ — original + bullet summary + page map
• index.md update — page catalog with descriptions
• log.md update — append-only ingest changelog

How to install

1. Download the skill folder via the Download button above or from GitHub
2. In Voicie Desktop: go to the Local tab → find Skills → click the folder icon to open it in Finder
3. Move the downloaded skill folder into that directory
4. Open a new chat — the skill is now available
5. Call it with natural language or /<skill-name>

→ Full guide: How to install skills in Voicie