Skill Library
An agent is as good as the procedure it is handed. Mine live in one repository of 59 skills, kept in step across three machines. Some I wrote from problems I solved once and did not want to solve again. 36 came from other people, and for each of those the commit and the licence are on record.
- 2026
- Agents
- Workflow
- Design system
In short
- Size
- 59 skills and 4 agents
- Borrowed
- 36 skills from 6 repositories, each with its commit and licence recorded, and 7 local edits written down
- Writing line
- Claude writes, DeepSeek rewrites, a script checks what must survive, Gemini reads and only flags
- Second opinion
- Asked for by rule: before handing over about 150 lines or 5 files, after two failed tries at one bug, before a decision that is hard to undo
- Design system library
- 16 sites measured with one fixed schema
- UX audit
- 7 sectors, 12 surfaces, every finding sourced
- Machines
- 3, synced through git. Credentials are never synced
A writing line across three models
Text written by one model reads like that model. So prose I deliver goes down a line. Claude writes it. DeepSeek rewrites it with one job: remove the patterns of machine writing, such as inflated importance, reflexive triplets and stock openings, while keeping every fact. A script then compares the two versions and lists any number, link or tag that went missing, because a rewriting model will drop them now and then. For translations a third model, Gemini, reads as a native speaker and is told to flag and not to rewrite.
None of the three publishes anything. The last step is a person reading it.
A second opinion, by rule
One skill hands work to another vendor's model for review, read only. It is not left to mood. The rules say when: before I hand over a change of about 150 lines or 5 files, when the same bug has survived two attempts, before a choice of architecture or library that is hard to reverse. The skill also says what to do with the answer: open the file and check every finding, because another model's opinion is not evidence and it can invent a line number.
Measuring other people's design systems
When a site is good I measure it instead of bookmarking it: the type scale, the grid, the spacing unit, the motion, all into one fixed schema, in passes that start with pinning the viewport. A field that cannot be measured is written as not measured, never filled with a likely value. With 16 sites in, a claim like most of these sites do this can be counted.
The library has a rule about originality that works by shape and not by warning. Method is free to reuse: a grid formula, a scale ratio, a token architecture. Expression is not: a palette, a type pairing, a signature interaction. So when the library produces a starter kit, those slots come out empty. A warning can be skipped. An empty slot cannot.
Knowing where things came from
36 of the skills were written by other people. They are copied in, which means updates do not arrive on their own, so a ledger records the source repository, the exact commit and the licence of each, along with the 7 edits I made and the skills I looked at and did not take. One source has no licence file at all, and the ledger says what that means if the repository is ever made public.
The library syncs across three machines through git. The session hook that notices a newer version only reads the remote commit and asks. It does not pull. If the repository were ever compromised, a person saying yes is the only defence left.