作者盤點 23,064 個智能體技能檔案後,構建 mstack 技能庫
I counted 23,064 agent skills before writing one. Here is what I built instead.
Manjunatha Sai Uppu 盤點 12 個公開技能庫中的 23,064 個 SKILL.md 檔案後,構建了 mstack 技能庫。mstack 分為 15 個 packs、共 115 項技能,要求技能執行腳本並回報 pass、fail、not verified 或 blocked,另以應觸發和不應觸發的請求測試技能描述。
Agent skills are folders with a SKILL.md file that teach a coding agent, such as Claude Code, Codex or Cursor, how to do one job. There are a lot of them now. I had about 720 skill files installed on my own machine.
Then I looked at my own history. Across several hundred coding sessions, only 26 different skills had ever started. The other 700 were sitting in the agent's context doing nothing, or worse, competing with the ones I needed.
So before building my own library, I counted what exists.
What 23,064 skill files look like
I downloaded twelve public skill libraries and wrote a script to read every SKILL.md: 23,064 files, 10,579 different skill names once the copies are removed.
Of those names:
- About 55% have a description that never says when to use the skill. An agent picks a skill from its name and description. If the description does not say when, the skill cannot be chosen at the right moment.
- 19% ship a script. The rest are text.
- 1% ship tests of their own.
Then I read a few hundred of them closely, domain by domain. The pattern was the same everywhere. Weak skills are lists of advice. Strong skills run a tool and end with a verdict. An advice skill cannot fail, so it cannot tell you anything.
The four rules of mstack
mstack is the library I built from that. Every skill follows four rules.
1. It runs a script. The script produces the check. The model does not grade its own work.
2. It ends in one of four states: pass, fail, not verified, or blocked. "Not verified" never counts as a pass. If a tool is missing, the skill says so and stops. It does not pretend.
3. It is tested for when it starts. Each skill has ten requests where it should start and ten where it should stay quiet. The quiet ones are near misses: requests that belong to the skill next door. The results are published in the README, including the bad ones.
4. Packs are small. A coding agent only has room for so many skill descriptions. In Claude Code the skill listing gets about 1% of the context window, and when it overflows, the least-used descriptions are dropped. So mstack is split into packs by domain. You install the pack for the work you do.
What the core pack does
You do not type commands. You say what you want.
| You say | Skill | What happens |
|---|---|---|
| "are we done?" | verify |
Runs your project's own test, lint and build commands and records each result against the exact code state. Edit a file afterwards and the record is marked stale. |
| "delete the old migrations folder" | safe-delete |
Lists what would go, which files still refer to it, and what git cannot bring back. Moves it to a trash folder. Gives you one undo command. |
| "only touch files in src/billing" | scope-lock |
Locks edits to that folder until you release it. |
| "we're going with Postgres, note that" | remember |
Saves a short decision record with its reason. Refuses a duplicate and refuses anything that looks like a password. |
| "I'm stopping for today" | handoff |
Writes one file a fresh session can continue from, and warns if the code has moved since. |
| "what did you just change?" | debrief |
Explains the real diff, and lists what is not done and not verified. |
| "is this skill safe to install?" | skill-check |
Reads a third-party skill without running it and reports what a person should look at. |
| "is there a skill for this?" | mstack |
Looks in the skill map and names one skill, or says none fits. |
The pack also ships a guard hook for Claude Code. It stops a recursive delete and points to safe-delete, and it asks before commands such as git reset --hard, a forced push or DROP TABLE. It also asks before MCP tools whose names look destructive, because a command filter cannot see what an API call deletes.
The test results, including the one that failed
Each core skill got ten "should start" requests and ten "should stay quiet" requests, in a clean session with only that pack installed.
| Skill | Started when it should | Started when it should not |
|---|---|---|
handoff |
10 of 10 | 0 of 10 |
mstack |
10 of 10 | 0 of 10 |
scope-lock |
10 of 10 | 0 of 10 |
skill-check |
10 of 10 | 0 of 10 |
debrief |
9 of 10 (7 before a fix) | 0 of 10 |
remember |
9 of 10 | 0 of 10 |
safe-delete |
9 of 10 | 0 of 10 |
verify |
8 of 10 | 0 of 10 |
Zero false starts in 80 tries. The first time I ran these, debrief started on only 7 of 10, below my pass mark of 8. I reworded its description using the three requests it had missed, ran the test again, and it reached 9.
The other 107 skills have a smaller smoke test so far: two fitting and two near-miss requests each. Across all 115 skills, they started on 277 of 294 fitting requests (94%) and on 1 of 294 requests meant for a neighbour. That one false start is listed in the README with the rest.
These are single runs. With many other skills installed, results will differ. The second kind of test, "is the result better with the skill than without it", exists for every skill and has not been run yet.
What my own tests caught
Writing the tests first paid for itself several times in one day:
-
rememberaccepted the same decision twice when it was reworded. Fixed. -
rememberdid not recognise a password written as "the password is ...". Fixed, in both wordings. - The first version of
skill-checkreported 17 failures in a well-known official skill pack. All 17 were false alarms: an API table containing the word "credentials", a safety note that quoted a dangerous command in order to forbid it. A scanner that cries wolf on good skills is useless, so I rewrote it to need real evidence. - My own results script recorded one score upside down. I caught it reading the output.
None of this would have shown up in a skill that only gives advice.
What is in the library
mstack has 15 packs and 115 skills: the core pack above, plus packs for software development, testing, requirements and brainstorming, design, media editing, writing, VR and AR, 3D assets, AI agents, operations, the web, data, apps, and growth.
The state of each skill is in the README table, generated from the skills themselves. Where a skill needs a tool I did not have, such as a 3D program or a headset, it detects the tool and stops as blocked, and its notes say that the live path is untested.
Numbers that go out of date, such as platform limits and thresholds, live in dated data files and are marked unverified until someone checks them against the source.
Install
Claude Code:
/plugin marketplace add manjunani/mstack
/plugin install mstack-core@mstack
Other agents that read the Agent Skills format can use the skill folders directly. They are Markdown and standard-library Python.
To choose how much the skills explain, add one line to your project's CLAUDE.md or AGENTS.md:
mstack-level: beginner
beginner, standard or expert. It changes the amount of explanation. It never removes a check.
What I would like from you
Try the core pack on a real project and tell me where a skill starts when it should not, or stays quiet when it should start. Those are the two failures I can measure and fix.
The method is in the repository under standard/: the rules, how to write a description that triggers, the template and how the tests work. A new skill is welcome when it does a job no existing skill does, and runs something.
Repository: https://github.com/manjunani/mstack
來源:DEV Community · dev.to