wiki-write-articles

name	wiki-write-articles
description	Multi-agent article generation from wiki data. Research agent queries sources, writer agent produces articles in target voice.

Wiki Write Articles

Generate wiki articles from corpus data using a two-agent workflow: a research agent that queries indexed chunks to gather evidence, and a writer agent that produces polished articles in a target voice with inline citations.

When to Use

After at least one search index is available:

• QMD local index (from wiki-index-qmd), OR
• Qdrant Cloud embeddings (from wiki-embed-qdrant)

Also requires entity inventory (from wiki-extract-entities) for topic selection and cross-referencing.

Prerequisites

• Search index: QMD collection or Qdrant embeddings populated
• Entity inventory at data/entities/entity-inventory.json and data/entities/top-entities.json
• Article topics selected (typically from top entities or SEO research)

Environment

Variable	Required	Source
QMD CLI	If using QMD search	Installed via bun
GEMINI_API_KEY	If using Qdrant backend	~/.zshrc
QDRANT_API_KEY	If using Qdrant backend	~/.zshrc
QDRANT_CLOUD_URL	If using Qdrant backend	~/.zshrc

Claude agents (research + writer) run within Claude Code sessions. No separate API keys needed for the agent layer.

Workflow Overview

Topic Selection ──► Research Agent ──► research.json ──► Writer Agent ──► draft-{voice}.md
                         │                                    │
                         ▼                                    ▼
                   QMD/Qdrant search              Voice guidelines + citations

Step 1: Topic Selection

Choose article topics from one or more sources:

From entity inventory:

# View top entities
python3 -c "
import json
top = json.load(open('data/entities/top-entities.json'))
for e in top[:20]:
    print(f\"{e['name']} ({e['category']}): {e['mentions']} mentions, {e['sources_count']} episodes\")
"

Good article candidates have:

• High mention count (discussed frequently)
• High source count (discussed across many episodes, not just one)
• Rich co-occurrence graph (connected to other entities)

From SEO research:

• Content briefs in data/seo/ (if generated)
• Keyword research targeting search volume for the wiki's domain

From editorial judgment:

• Core concepts the show is known for
• Frequently asked questions from the audience
• Controversial or distinctive positions taken by hosts

Step 2: Research Agent

The research agent gathers evidence for a specific topic by querying the search index.

Research queries (3-5 per topic):

1. Direct entity name search
2. Related concept search
3. Specific quote/advice search
4. Contrarian or nuanced angle search

Using QMD:

qmd search --collection {slug}-transcripts --query "Alex Hormozi business advice" --limit 10
qmd search --collection {slug}-transcripts --query "Hormozi $100M Offers framework" --limit 5
qmd search --collection {slug}-transcripts --query "Hormozi controversial opinion" --limit 5

Using Qdrant backend:

curl -X POST https://api-production-4224.up.railway.app/chat \
  -H "Content-Type: application/json" \
  -H "X-User-Email: [e-mail]" \
  -d '{"wiki":"{slug}","query":"Alex Hormozi business advice","limit":10}'

Fallback (local chunk search): When API limits are exhausted, search chunk files directly:

grep -rl "Hormozi" data/chunks/ | head -10

Research output: Save as docs/articles/{topic-slug}/research.json:

{
  "topic": "Alex Hormozi",
  "category": "people",
  "queries": [
    {"query": "Alex Hormozi business advice", "results": 10}
  ],
  "quotes": [
    {
      "text": "Direct quote from transcript...",
      "episode_id": "episode_slug",
      "episode_title": "Episode Title",
      "timestamp": "15:30",
      "speaker": "Sam"
    }
  ],
  "key_insights": [
    "Insight summary with episode reference"
  ],
  "episodes_referenced"   
     

Step 3: Writer Agent

The writer agent produces the article from the research file, following voice guidelines.

Input: research.json for the topic Output: draft-{voice}.md (e.g., draft-housel.md)

Voice Selection

Choose a voice style for the wiki. The voice should match the wiki's audience and tone.

Proven voices from existing wikis:

Morgan Housel voice (used for MFM Wiki):

• Calm, measured tone. Zero exclamation points.
• Lead with counterintuitive observation or striking anecdote
• Concrete numbers and specific examples throughout
• Paraphrase by default; only use direct quotes when "clip-worthy gold"
• Inline YouTube links with timestamps
• Wiki links using [[Entity Name]] for cross-references

Other voice options to consider:

• Packy McCormick: Energetic, pop-culture analogies, punchy subheads
• Encyclopedia: Neutral, factual, Wikipedia-style
• Practitioner: How-to focused, step-by-step, actionable

Article Structure

# {Entity/Topic Name}

Opening paragraph: counterintuitive hook or striking fact from the data.

## Background / Context

Who/what this is, why it matters in the show's universe.

## Key Themes

### Theme 1: [Specific Angle]

Content with inline citations: "direct quote" ([Episode Title, MM:SS](youtube_url&t=XXs)).

### Theme 2: [Another Angle]

More content with cross-references to [[Related Entity]].

## Notable Quotes

> "Exact quote from transcript" -- Speaker Name, [Episode Title](url)

## Episodes

| Episode | Date | Key Topics |
|---------|------|------------|
| Episode Title | YYYY-MM-DD | topic1, topic2 |

## See Also

- [[Related Entity 1]]
- [[Related Entity 2]]

Step 4: Batch Processing

For producing multiple articles in a batch, use a manifest:

docs/articles/BATCH-MANIFEST.json:

{
  "batch_id": "batch-2026-02-12",
  "status": "in_progress",
  "voice": "housel",
  "articles": [
    {"id": "alex-hormozi", "status": "draft_complete"},
    {"id": "boring-businesses", "status": "researching"},
    {"id": "network-effects", "status": "todo"}
  ]
}

Process each article sequentially:

1. Create docs/articles/{topic-slug}/ directory
2. Run research queries, save research.json
3. Write draft, save draft-{voice}.md
4. Update manifest status

Output

Per Article

docs/articles/{topic-slug}/
├── research.json       # Gathered evidence with citations
└── draft-housel.md     # Article in target voice

Batch

docs/articles/
├── BATCH-MANIFEST.json
├── alex-hormozi/
│   ├── research.json
│   └── draft-housel.md
├── boring-businesses/
│   ├── research.json
│   └── draft-housel.md
└── ...

Quality Checks

Research Quality

• 3+ queries per topic: Each topic should have multiple search angles
• 5+ source episodes: Articles grounded in a single episode are thin
• Direct quotes captured: At least 3-5 clip-worthy quotes with timestamps
• Cross-references identified: Related entities noted for wiki linking

Article Quality

• Factual accuracy: Every claim should trace back to a source chunk. No hallucinated quotes, anecdotes, or statistics.
• Voice consistency: Read the article aloud. Does it match the target voice? (e.g., for Housel: no exclamation points, no listicles, calm tone)
• Inline citations: Quotes and specific claims have [Episode Title, MM:SS] links
• Wiki links: Related entities use [[Entity Name]] linking
• No fabricated content: Everything must come from the research.json or be clearly labeled as editorial context
• Length: Wiki articles should be 1,000-3,000 words. Shorter for minor entities, longer for major topics.
• Episode table: Includes a table of referenced episodes

Common Failure Modes

• Hallucinated quotes: The writer agent invents quotes that sound plausible but are not in the research. Always cross-check direct quotes against research.json.
• Wrong attribution: Quote attributed to wrong speaker. Verify speaker labels from source chunks.
• VIDEO_ID placeholders: YouTube links need real video IDs. These may need a separate episode-to-YouTube mapping step.
• Over-reliance on one episode: Good articles draw from 5+ episodes. If research only found 1-2 episodes, the topic may not warrant a standalone article.

Shared Library / Resource Reference

Resource	Location	Purpose
Entity inventory	data/entities/top-entities.json	Topic selection
Chunk files	data/chunks/*.json	Fallback local search
Shared backend	wiki-projects/shared-backend/	RAG API
MFM articles	wiki-projects/mfm-wiki/docs/articles/	Reference implementation (49 articles)
Batch manifest	docs/articles/BATCH-MANIFEST.json	Batch tracking

Troubleshooting

RAG API returns no results: Check that the Qdrant collection has vectors (/stats/{slug} endpoint). If empty, run wiki-embed-qdrant first.

Daily query limit exhausted: The backend has a 20-query daily limit per email. Use a different email or fall back to local chunk file search with grep.

Articles feel generic: Increase research depth. More queries with more specific angles produce more distinctive articles. Search for the entity's name plus specific subtopics, controversies, or frameworks.

Missing YouTube URLs: Create an episode-to-YouTube mapping file. This is a manual or semi-automated step that matches episode IDs/titles to YouTube video URLs.

Batch consistency: When writing 10+ articles in one session, the writer agent may drift from the voice guidelines. Re-read the voice spec every 5 articles and compare against the template article.