X Account Enrichment

Enrich skeleton entity pages for X/Twitter accounts tracked in ~/x-accounts.yaml.

Entity page names may differ from the handle or display name

• Before creating/updating an entity page for an account, grep -ril "<person name or handle>" wiki/entities/ — pages are titled by display name (xeophon → florian-brand.md, _lopopolo → ryan-lopopolo.md), and a handle-grep alone misses them.
• A person may have TWO pages: a person page and a blog/writings page (ryan-lopopolo.md vs ryan-lopopolo--writings.md). Article-analysis content belongs on the writings page; tweet-level activity on the person page. Check both before creating anything new.

Workflow

⚠️ CRITICAL: Git History Check First
Before enriching ANY page, check if a richer version exists in git history. Use bash config/hermes/skills/wiki/wiki-entity-enrichment-from-article/references/find-richest-version.sh wiki/entities/<handle>.md 3. If the richest version has >50 more lines than current, restore it first, then enrich on top. See wiki-entity-enrichment-from-article → references/pre-write-verification.md → "Git History Enrichment" for the full pattern.

1. Audit current state:

   python3 -c "
   import os, yaml
   with open(os.path.expanduser('~/x-accounts.yaml')) as f:
       accounts = [a['handle'].lstrip('@') for a in yaml.safe_load(f)['accounts']]
   entity_dir = os.path.expanduser('~/wiki/entities')
   for handle in accounts:
       path = os.path.join(entity_dir, f'{handle}.md')
       if os.path.exists(path):
           with open(path) as f: content = f.read()
           skeleton = 'status: skeleton' in content
           enriched = 'Core Ideas' in content
           print(f'  {handle}: skeleton={skeleton}, enriched={enriched}')
       else:
           print(f'  {handle}: MISSING')
   "

2. Prioritize by tiers:

   • Tier 1: High-impact AI researchers, open-source leaders, known bloggers
   • Tier 2: ML engineers, startup founders, educators
   • Tier 3: Contributors, emerging voices

Batch process 2 per subagent (4 parallel max):

• Split entities across 4 parallel subagents, 2 per subagent
• Provide exact filenames in the goal (absolute paths: /home/exedev/wiki/entities/{name}.md)
• Include the format template in the prompt
• 8 entities processed this way = ~6 minutes total (vs 15+ minutes sequential)

4. Post-batch verification:

   # Check file sizes
   ls -la ~/wiki/entities/*.md | sort -k5 -n -r | head -30
   
   # Find remaining skeletons
   grep -l 'status: skeleton' ~/wiki/entities/*.md
   
   # Find duplicates (skeleton + enriched for same person)
   # Common pattern: full-name-slug.md vs handle.md

5. Cleanup duplicates:

   # Delete skeleton duplicates after confirming enriched version exists
   rm ~/wiki/entities/ethan-mollick.md      # → emollick.md
   rm ~/wiki/entities/chip-huyen.md          # → chipro.md
   rm ~/wiki/entities/lilian-weng.md         # → lilianweng.md
   rm ~/wiki/entities/eugene-yan.md          # → eugeneyan.md
   rm ~/wiki/entities/samuel-colvin.md       # → samuelcolvin.md (also empty)
   rm ~/wiki/entities/benjamin-clavi.md      # → bclavie.md
   rm ~/wiki/entities/cl-mentine-fourrier.md # → clefourrier.md
   rm ~/wiki/entities/late-interaction.md    # concept, not person

6. Commit and push:

   cd ~/ai-topics && git add wiki/ && git commit -m "wiki: enrich X accounts (batch N)" && git push

Enrichment Format Template

---
title: "Full Name"
handle: "@twitter_handle"
created: 2026-04-10
updated: 2026-04-10
tags: [person, topic1, topic2]
aliases: ["handle", "alt-name"]
---

# Full Name (@handle)

| | |
|---|---|
| **X** | [@handle](https://x.com/handle) |
| **Blog** | [URL](URL) |
| **GitHub** | [username](https://github.com/username) |
| **Role** | Job title |
| **Known for** | Key contributions |
| **Bio** | 2-3 sentence background |

## Overview

2-3 paragraph introduction of who they are, their background, and why they matter in AI.

## Core Ideas

Their key viewpoints, theories, and opinions on LLM/AI Agent technologies. Use subsections for each major theme. Quote their actual posts/articles where possible.

## Key Work

- Project/tool/library they created
- Papers published
- Notable blog posts
- Talks and presentations

## Blog / Recent Posts

Key articles they've written with dates and summaries.

## Related People

Connections to other wiki entities.

## X Activity Themes

What they tweet about most frequently.

Quality Targets

• Size: 8-15KB per page minimum
• Content: Actual quotes, specific examples, real blog post links
• Sections: Core Ideas (with subsections), Key Work, Blog/Recent Posts, Related People, X Activity Themes
• Frontmatter: No status: skeleton tag
• Cross-references: Link to other wiki entities using [[entity-name]] format

Wiki commit conventions in ai-topics (pitfalls — learned 2026-09-19)

1. Tag taxonomy pre-commit hook: Every frontmatter tag in new/edited wiki pages must exist in wiki/SCHEMA.md Domain Concepts line (~line 40, 950+ tags). Unknown tags BLOCK the commit. Either map to an existing canonical tag (grep SCHEMA.md first — it already has ai-company, startup, llm, etc.) or append the new tag to the Domain Concepts line. Note SCHEMA.md lives at wiki/SCHEMA.md (run python from inside wiki/).
2. Japanese-in-clean-files hook: New wiki entity/concept files must contain ZERO Japanese characters — even inside wikilink section refs like [[page#日本語|label]] triggers "NEW FILE with Japanese content" and blocks the commit. Write all new wiki pages in English.
3. Index sync: wiki/index.md is regenerated by a daily cron — don't edit it manually, it causes rebase conflicts with other concurrent jobs.

Related References

• references/scan-naming-conventions.md — Filename conventions for scan-sourced pages
• references/tweet-content-extraction.md — Extracting full content from tweets
• references/x-accounts-scan-json-structure.md — JSON structure of x_accounts_latest_full.json
• references/x-scan-discord-report-template.md — Japanese Discord report format template for x-accounts-scan cron job

Known Subagent Pitfalls

1. Filename aliasing: Subagents create files with different names than specified (e.g., hynek-schlawack.md instead of hynek.md). Always audit post-batch for duplicates.

2. Budget exhaustion: 50-iteration budget per subagent. When processing 5 entities, subagent may hit limit and skip writing some files. Check exit_reason: max_iterations in results.

3. Path discipline: delegate_task subagents must always target ~/wiki/entities/. If a delegated context shows ~/.hermes/home/..., treat it as a runtime artifact and keep writing to the canonical wiki path only. grep -ril output showing /opt/data/.hermes/home/wiki/... paths is the same symlink artifact — those ARE the canonical ~/wiki files, not a second copy.

4. Status cleanup needed: Even when subagents write content successfully, status: skeleton in frontmatter is often NOT removed. Always verify and manually replace with status: complete if needed.

5. Successful pattern confirmed (2026-04-10): 8 entities enriched in 4 parallel batches of 2 each, completed in ~6 minutes with file sizes 12.5-18.7KB (exceeding 8-15KB target). This confirms the 2-per-subagent batching strategy works well.

6. Handle ≠ filename during scan ingest: When processing scan results, the X handle (e.g., dbreunig) often does NOT match the entity page filename (e.g., drew-breunig.md). Always search for the entity page by content/grep before concluding it's MISSING:

   grep -rl 'dbreunig' ~/wiki/entities/ 2>/dev/null

Research Strategy

For each person:

1. Search X/Twitter for their recent activity and key opinions
2. Check their blog/personal site for articles
3. Look up GitHub repositories they've created
4. Find interviews, talks, or presentations
5. Search for mentions in AI community discussions
6. Cross-reference with other wiki entities for related connections

x-accounts Scan → Wiki Ingest (different workflow)

This skill covers enriching skeleton entity pages for tracked X accounts. A separate workflow exists for processing new posts from x-accounts scans into wiki pages (creating concept pages from shared links, updating entity pages with new blog posts, etc.).

For the scan→ingest workflow, see wiki-ingestion-pipelines → references/x-accounts-scan-ingest-workflow.md.

Key distinction:

• This skill: Take a status: skeleton entity page → research the person → write 8-15KB full page
• Scan ingest: Take new_posts from x_accounts_latest_full.json → evaluate each post's external URLs → create concept pages or update existing entity pages

Post evaluation quick reference:

Signal	Action
Reply without standalone value	Skip
Multiple replies from same account forming coherent thread (same topic, consecutive timestamps)	Evaluate as group — may have standalone value together
Non-AI domain link	Skip
Link to own blog post (entity exists)	Update entity's Blog/Recent Posts
arXiv paper (no existing page)	Create concept page
Official docs link	Add to relevant entity sources
Conference appearance	Add to entity's Speaking Engagements