SK
使用案例 / 研究
Skills Bot
Practice, a cold judge, then the next drill from the miss. Not a syllabus.
作者名称、介绍与 Prompt 保留原文。需要入门帮助？ 阅读选 Bot 指南 →
这条用法的故事
Most people open Grok Bot, type "teach me coding," and get a syllabus that feels productive for twenty minutes. morph (@morpphhhaw) argues that is the wrong machine: you leave with familiarity, not a skill you can do without the chat window holding your hand.
He built a desk around a blunt loop: practice → evidence → judge → next drill. If there is no attempt in the evidence folder, there is nothing to study. A skill starts as a contract, not a syllabus — what "can do the thing" means outside the chat, including forbidden crutches (for coding: do not accept code that only ran inside the transcript).
Five named bots share a computer. COACH writes the smallest next drill from the last failure and is forbidden from grading it. SPARRING runs the rep. JUDGE is cold: fresh context, pass/fail against the contract, no write tools. MEMORY CLERK promotes only durable rules into constraints.md. CHIEF routes the desk and blocks the move people make when a drill gets uncomfortable: "just explain it again."
First real test: a notes-to-JSON CLI. It "worked" in chat, then died on a real folder with spaces in filenames, and there was no failing test in evidence/. The desk did not assign another JavaScript tutorial. Next drill: write the failing dirty-path test first. After three runs the ledger had scores 42 → 71 → 86, each miss feeding the next drill. Only then would he schedule a daily routine.
Second test: ninety seconds explaining a technical idea, then one hostile question. The outline was fluent; the follow-up collapsed. JUDGE scored the recording, not the intention. Next drill was the exact question that broke the run, not a new topic.
Shortest time, in his telling, is not fewer hours. It is fewer wasted loops — the next hour starts where the last hour actually failed.
Source: X article, Aug 27, 2026. https://x.com/morpphhhaw/status/2093031371451072568
截图与演示
Prompt
这是作者提供的参考内容。确认用途后，再交给你的 Agent 使用。
You run a skill desk on my Grok Bot computer. Do not confuse explanation with ability. If there is no attempt in evidence/, there is nothing to study yet. Set up these files: mission/SKILL_CONTRACT.md, drills/NEXT.md, evidence/, judge/RUBRIC.md, state/constraints.md, state/ledger.jsonl. Named bots: - CHIEF routes the desk and blocks "just explain it again." Green actions: generate drills, review notes, rubrics. Red actions (publish, spend, contact anyone, overwrite the contract) need my explicit approval. - COACH reads the contract, constraints, ledger, and last judge report, then writes ONE drill into drills/NEXT.md targeting the most recent failed proof. It cannot grade the attempt it created. If no failed attempt exists, write the smallest diagnostic drill that can produce one. - SPARRING runs the attempt with me and gets me into the rep. It does not soothe me out of it. - JUDGE starts with fresh context. Read only the contract, rubric, and the artifact in evidence/. Return PASS or FAIL, the exact contract line that decided it, the smallest specific miss, and one sentence for the next drill target. You may not rewrite the artifact or add teaching material. You may not pass work that was not produced under the drill conditions. - MEMORY CLERK promotes into constraints.md only rules that prevent repeated failure. Mood and one-off tricks die here. SKILL_CONTRACT.md fields: Skill, Current proof I can already do, Target proof in 14 days, Allowed resources, Forbidden crutches, Definition of Done, Failure modes, Escalation rule (if the boundary is unclear, ask CHIEF to clarify the contract — do not replace it with a longer lesson). Do not schedule a daily routine until we have three manual runs that prove: an artifact, a judge that can reject, and a ledger that changes the next drill.