# The learning loop (every role, every run)

Agents do not get better between runs on their own. Each run starts from zero, so whatever a role learned has to be written down and read back. This is the loop.

1. BEFORE: read the top LESSONS block of your NOTES.md (at most ~15 bullets, the distilled rules) and any "SCORECARD SAYS" line the numbers agent left. Apply them this run.
2. DURING: treat each action as a small experiment. Know which variable you are testing (hook shape, venue, time, length, format, offer).
3. AFTER: append one entry to NOTES.md:

```
TRIED: <what you did + the variable you changed>
RESULT: <real numbers, or "pending, check at HH:MM">
LESSON: <keep / drop / try next>
```

4. CHECK BACK: at the start of each run, fill in the pending results from your previous run. A lesson without a measured result is a guess; mark it "guess".
5. DISTILL: when a lesson is confirmed twice, promote it into the LESSONS block at the top of NOTES.md. Remove lessons the data has disproved. The numbers agent prunes and merges these.
6. SHARE: a lesson that applies to other roles goes into LEARNINGS.md as one line: date, source role, evidence.

Judge every number only against your own history (1 day ago, 7 days ago, personal best, this post vs the median of your last 10). A flop is the content's fault: diagnose the post, never blame follower count or timing.