I want a skill for making skills. I already have three, and the size spread between my skills is 267 to one.
I will run a deep research into how other teams and people set up skill creation.
I want a separate skill for making skills, with the methodology inside it: how to design them properly. Skills should be ergonomic, high quality, maximally clear and understandable, without unnecessary complexity and without entropy growth. Understandable enough that anybody should be able to use them, even a very unprepared person.
I need to work out which types of skills are better to create, by what logic to design them, and how broad or detailed the rules and descriptions should be. The outcome I want is one clear algorithm by which all my skills get created.
I ran the recall. It closed half of this with three existing skills, and the other half with a measurement that explains exactly what I have been feeling.
⭐⭐ Three skills for making skills are already on the shelf. skillify at 46KB, last touched 22 September, the largest of the three. skill-gap at 12KB, 2 October, which finds the hole where a skill is missing. skill-forge at 5.9KB, 2 October, the smith itself. Plus skill-creator beside them as a plugin.
So "I want a separate skill for making skills" describes a shelf that already holds three of them, built across two months, and not one of them absorbed the others.
The shelf itself, in numbers. 282 directories, 277 with a SKILL.md, 3,683KB in total. That is 3.6 megabytes of instructions for one operator.
⭐⭐⭐ And here is the number that names what I am actually asking about. The median skill body is 9,315 bytes. The largest, dr-fanout, is 179,104 bytes. The smallest, screenpipe-worklog, is 671 bytes.
A spread of 267 to one, and nowhere is there a single written rule about how large a skill should be.
That is "entropy growth" with a figure attached. My own question about how broad or detailed the rules should be currently has no answer at all: the top five are dr-fanout 174KB, apply 125KB, tt 122KB, retro 94KB, rep-reply 52KB, against a median of nine kilobytes and a floor under one.
⚠️ But the discipline I want does exist for one half, and it holds at 96 percent. I have a standing rule: a skill description is at most 400 characters, on one quoted line. And there is a live instrument enforcing it, skill_desc_audit.py at 12KB.
Measured across all 277: zero skills with no description, and 13 over 400 characters, which is 4 percent. The worst are local-to-call at 1,281 characters, local-brain-onboard at 1,133, local-fb-messenger at 1,003. All 13 are local skills rather than fleet-wide ones.
⭐ So the gap is precise rather than general. Where a rule and a door exist, compliance is 96 percent. Where no rule exists, the spread is 267-fold. The methodology I am asking for is needed for exactly one dimension, the body, and it is needed because nobody ever wrote a sentence about it.
Two more instruments already stand beside the shelf. skill_usage_log.py at 13KB counts whether a skill was actually USED rather than merely invoked. And codex_skills_bridge.py at 18KB mounts the same file for a different agent so no second copy is made.
Which means my "one clear algorithm" already has three of its four parts: a creator, a gap-finder, a description auditor, a usage counter and a cross-agent bridge. What is missing is one page stating how large a skill may be, what it must contain, and when to split it instead of growing it.
⚠️ And one thing worth doing before commissioning research into how other people work. A 3.6MB shelf of 277 skills is itself a standing context tax on every session. My question "which types of skills are better to create" has a local answer available today: run the usage counter over 30 days and see which of the 277 fired at all. Types that never fire do not need a design standard, they need deleting. That measurement costs nothing, and it sharpens the external research rather than replacing it.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261106.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261106.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)