Most "content idea" tools hand you other people's topics. This one starts from your own material: what you shipped, what you bookmarked, what came up on calls. Then it ranks ideas against your past winners.
# once: pick your past winners (needs posts with numbers, see voice/)
python gather/make_winners.py --top 40
# every morning
python gather/bookmarks.py --user-id <your numeric id> # new X bookmarks
GITHUB_OWNERS="you your-org" gather/github_activity.sh 24 > gather/data/github.md
# ...drop any call notes into gather/inbox/
python gather/build_context.py # -> gather/data/context-<date>.md
# give your AI assistant gather/prompts/daily-brief.md + the context file
# -> gather/data/brief-<date>.md
# rank the ideas
LLM_CMD="claude -p" python gather/score_ideas.py gather/data/brief-<date>.md| File | What it does | Needs |
|---|---|---|
bookmarks.py |
New X bookmarks since last run, as a short digest | X_OAUTH2_USER_TOKEN (bookmark.read) |
github_activity.sh |
PRs and commits from the last N hours across your repos | gh CLI, logged in |
inbox/ |
Anything you drop in: call notes, transcripts, voice memos | nothing |
build_context.py |
Bundles all of the above + VOICE.md + winners into one file | nothing |
prompts/daily-brief.md |
Turns the context into 3-7 ideas with hook, beats, risk and source | any AI assistant |
score_ideas.py + scoring/rubric.md |
Scores each idea 0-100 vs your winners, clamps the math | any CLI model via LLM_CMD, or paste the prompt |
make_winners.py |
Picks your best past posts from voice/data/posts.jsonl |
posts with engagement numbers |
Everything writes to gather/data/ and gather/inbox/, both git-ignored. Run it by hand, from cron, or as a scheduled task in your AI assistant.
The brief prompt has three rules worth keeping even if you change everything else: never invent a number, keep private things private, and say what you held back.