fact-rotationlisted
Install: claude install-skill Tommyapi/ai-employee-skills
# Fact rotation
An AI writer left alone reuses its one or two favourite stories. In our own audit, one story filled half of the output. The fix is mechanical: facts live in one file, a usage log records every use, and the picker hands the writer the least-used ones.
## When to use
- Before drafting: ask for N facts instead of letting the model choose.
- After publishing: log the facts you actually used (the log is what makes rotation work).
## Setup
1. Keep a `FACTS.md` with one fact per line:
`- **F12** [theme] fact text | fit: blog,x | limit: do not name customers`
`fit` is a platform list or `all`; `limit` is a free-text restriction the writer must respect.
2. Keep a usage CSV next to it (the script creates it).
## Steps
1. Pick: `python scripts/facts_rotate.py --facts FACTS.md --usage usage.csv pick --n 5 --platform blog [--theme hiring]`
2. Write the draft using only those facts. Respect each fact's `limit`.
3. Log: `python scripts/facts_rotate.py --facts FACTS.md --usage usage.csv used F12,F30 --where blog:post-7`
4. Review rotation health any time: `... stats` (7-day and 30-day counts per fact).
## Rules baked in
- Least used in the last 7 days comes first; ties go to the oldest last use.
- A fact already used twice today is excluded from picks (`used` still records it and prints a WARNING, because the log must stay true).
- Unknown fact IDs are rejected, so typos never pollute the log.
## Pair with
`draft-fact-lint` to prove the finished draft contains no