| name | follow-builders |
| description | Pull the latest AI builder feeds (tweets + podcasts) from follow-builders into the knowledge base raw/ directory, then compile new content into wiki articles. Trigger when user says /follow-builders, "pull builder feeds", "fetch builder feeds", or "pull the latest AI posts". |
Follow Builders Skill
Pulls curated content from 25 AI builders on X and 6 podcasts into the local knowledge base.
Source: https://github.com/zarazhangrui/follow-builders (no API keys required).
Trigger
User invokes /follow-builders or says something like:
- "pull the latest builder feeds"
- "fetch builder feeds"
- "pull AI builder posts"
- "get the latest from follow-builders"
Workflow
Run these steps in order:
Step 1 — Build the ingest command
Parse the user's message for options:
- If user mentions a likes threshold (e.g. "min 200 likes", "high signal only") → add
--min-likes N
- If user says "dry run" or "preview" → add
--dry-run
- Default:
--min-likes 50
Step 2 — Run ingest_feed.py
python tools/ingest_feed.py [--min-likes N] [--dry-run]
This script:
- Fetches
feed-x.json and feed-podcasts.json from the follow-builders GitHub repo
- Skips already-ingested content (tracked in
tools/feed-state.json)
- Writes new tweets to
raw/ai-builders/x/YYYY-MM-DD-{handle}.md
- Writes new podcast episodes to
raw/ai-builders/podcasts/YYYY-MM-DD-{slug}.md
- Updates
raw/_sources.md with pending entries
If the script fails (network error, missing dependencies), report the error clearly and suggest the user check their internet connection or run pip install if modules are missing.
Step 3 — Report results
After the script runs, summarize:
- How many new X post files were written (and which builders)
- How many new podcast episodes were written
- If nothing new: say so clearly — "No new content since last run"
- If dry run: show the preview and ask if user wants to proceed
Step 4 — Offer to compile
If new files were written (not dry run), ask:
"N new files added to raw/ai-builders/. Compile them into wiki articles now?"
If user confirms (or already said "and compile" in the original request):
- Run the standard ingest & compile workflow from CLAUDE.md § Ingest & Compile
- Read each new raw file, identify concepts, create or update wiki articles
- Update
wiki/_index.md, raw/_sources.md (mark processed), wiki/_meta/changelog.md
Output Format
## Follow Builders — YYYY-MM-DD
**Fetched:** N tweets from M builders · P podcast episodes
### New X Posts
- Andrej Karpathy (3 posts) → raw/ai-builders/x/2026-04-03-karpathy.md
- Aaron Levie (1 post) → raw/ai-builders/x/2026-04-03-levie.md
### New Podcasts
- Training Data: How Autonomous Labs... → raw/ai-builders/podcasts/2026-03-24-...md
### Skipped (already ingested)
- 18 tweets already in state
---
Ready to compile into wiki articles? [yes/no]
Rules
- Never fabricate content — only process what the JSON feeds return. If a feed fails to load, skip it and report the error.
- Respect the state file — never reset or manually edit
tools/feed-state.json. It's the dedup guard.
- Grouped by builder per day — tweets from the same builder on the same day go into one file. Do not split per tweet.
- Compilation follows CLAUDE.md — when compiling, follow the full article template and index update rules from the project's CLAUDE.md.
- Always report what was skipped — tell the user how many items were filtered (already ingested, below min-likes threshold).
