X SkillOpt learn
use this when running the X SkillOpt learn loop: log every post trajectory, refresh metrics, and nightly gated-edit best_skill.md so the bot only doubles down on what works
YannisKiefer · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add YannisKiefer/grokbot-x --skill X SkillOpt learn技能说明
name: X SkillOpt learn description: >- use this when running the X SkillOpt learn loop: log every post trajectory, refresh metrics, and nightly gated-edit best_skill.md so the bot only doubles down on what works
X SkillOpt learn loop
Use this whenever shipping X posts/replies for the operator, and on the
nightly sleep routine. Treat skillopt/best_skill.md as the trainable skill
(Microsoft SkillOpt method adapted to X).
Decide, then ship
Do not ask the operator for strategy permission. Report live URLs or real blockers only. No "want me to" questions.
After EVERY ship (rollout)
- Read
skillopt/best_skill.mdBEFORE drafting (during a live trial on a trial day:skillopt/proposed.md). - Ship via the locked Typefully paths in
docs/PUBLISH.md. - Log it:
python3 skillopt/sleep_cycle.py append --json '{...}'with id, ts, type, url, text, features (viral_ratio, viral_signal, topic, author_followers = YOUR own follower count now, source), metrics placeholder, split (train|val ~80/20), notes. Features are frozen at ship time; later refreshes touchmetricsonly. On a trial day, add--skill skillopt/proposed.mdand use split val. - Refresh metrics later at +1h / +6h / +24h / next night
(
views/likes/replies/bookmarks).
Nightly sleep (reflect -> edit -> gate)
- Refresh metrics on posts older than 1h (read-only).
python3 skillopt/sleep_cycle.py status- If
ready_for_editis false (scored rows <min_trajectories_before_edit): log an epoch skip. Do not edit. - Else: write <= 4 bounded add/delete/replace edits into a copy,
skillopt/proposed.md, from train successes vs failures +rejected_edits.md. Targeting rules go in the skillopt-policy block. - GATE:
python3 skillopt/sleep_cycle.py gate accept --summary "..." --edit "..." --skill skillopt/proposed.md. The gate decides on held-out rows (live trial if logged, else replay) and adopts only if the gain's interval is above zero. Every look since the last accept (gate or replay; a new proposal, or the same one on new data) raises the bar, so try few, well-reasoned edits and gate a trial only once. Prose-only edits need a live trial first. Never editbest_skill.mdby hand. - Mirror accepted rules into the next day's hourly behavior.
- Stay quiet unless a skill edit was accepted (one line) or a hard blocker.
Live trial (prose edits)
- Start: copy
best_skill.mdtoskillopt/proposed.mdand edit it. - Alternate days: odd days draft from
proposed.mdand log with--skill skillopt/proposed.md; even days draft frombest_skill.mdand log normally. All trial rows use split val. - Stop: once both sides have >=
min_val_rowsscored rows, run the gate with--skill skillopt/proposed.md. Accepted or not, the trial ends.
Score
Fixed in skillopt/sleep_cycle.py; see docs/SKILLOPT.md. Cold account: relative scores matter more than absolute
views until you have a real ICP graph.
Reference
https://github.com/microsoft/SkillOpt Rollout, reflect, bounded edits, held-out gate, rejected buffer, slow update.