| name | x-ai-feed |
| description | Build a self-contained local HTML feed of recent posts from curated high-signal X accounts, for staying current on AI news/papers/projects without doom-scrolling. Use when asked to refresh the AI feed, check what's new on X, or update the news dispatch. |
X AI Feed
Builds ~/x-ai-feed/feed.html: a standalone page listing recent posts from a
curated list of X handles, grouped by category, newest first. No X API
account or MCP server required — it scrapes public profile pages via the
firecrawl CLI (check authentication with
firecrawl --status).
This trades the official X API's reliability for zero recurring cost: no
X Developer Portal signup, no $200+/mo API tier. The tradeoff is that it
depends on Firecrawl successfully rendering x.com profile pages, which can
change or get rate-limited without notice — if a run comes back mostly
empty or firecrawl --status shows depleted credits, that's the first thing
to check.
Running it
~/.agents/skills/x-ai-feed/scripts/feed.py
Flags (all optional):
--lookback-hours N (default 24) — only include posts newer than this--max-stories N (default 60) — cap on total stories in the feed--sources PATH (default sources.json next to this skill)--output PATH (default ~/x-ai-feed/feed.html)
Each run costs one Firecrawl credit per handle in sources.json — check
firecrawl --status before widening the source list or tightening the
schedule.
Workflow
- Confirm
sources.json exists and has handles the user actually wants
tracked. If the user asks to add/remove/recategorize accounts, edit
sources.json directly (it's a small hand-editable list, not generated).
- Run
scripts/feed.py with whatever --lookback-hours fits the ask (a
daily catch-up vs. "what happened since yesterday").
- Report back: story count, source count, and which handles (if any) came
back empty (
errors in the printed output / meta.errors in the page) —
don't just say "done," name what's missing so the user can judge if a
rerun or a sources.json edit is warranted.
- Open
~/x-ai-feed/feed.html (or point the user to it) rather than
summarizing its contents in chat — the whole point is a page they browse
themselves.
Editing sources
sources.json has a flat categories list and a sources list of
{handle, category} pairs (handle without @). Categories in sources.json
drive both the on-page filter tabs and the badge on each entry — keep the
category set small (4-6) or the tab row gets unwieldy.
Scheduling
For recurring runs (e.g. every 4 hours), use the current harness's automation
or scheduling facility to run the command above on an interval. Set this up
only if the user asks for it, since it is a standing automation, not a one-off.
Design notes for template.html
assets/template.html is a static template with two replacement tokens,
__FEED_DATA_JSON__ and __META_JSON__, filled in by feed.py via plain
string substitution (no templating engine). If the visual design needs
changing, edit the template directly — the data contract (Post fields in
feed.py: handle, author, verified, category, posted, url, text, likes,
retweets) is the thing not to break.
