two libraries, one craft
20 skills for strong readers, plus 41 for weak ones â including 11 that only make sense when you run models locally.
the songbook
Repertoire
A library of 61 skills that transfer engineering judgment to readers who don't have a senior to ask â engineers, and AI models of varying strength. Written like sheet music: play it as written, and it works.
61 skills
Markdown
two audiences
verified examples
A skill is a short, self-contained document that teaches one slice of engineering judgment â safe data handling, API design, git recovery, running local models â well enough that a reader can act on it without a senior to check with. This library has 61 of them, written for two very different audiences.
The first library targets strong readers: mid-level engineers and capable AI models. The second rewrites everything from scratch for weak readers â 7B-to-70B local models, often quantized, and first-time users â where a critical instruction buried past a thousand tokens simply never gets read, so the one load-bearing rule goes in bold, right under the title.
What makes it trustworthy is the process: every skill went through author, adversarial review, naive-reader review, and verification. Every runnable example was actually executed and the output shown is what really printed; local-model claims were tested against a live 7.5B model; anything only desk-checked is labeled as such.
highlights
verified, not vibed
Every number and output block comes from running the shown code. Fabricated example output was a fireable offense in review.
adversarial pipeline
Author â adversarial review â naive-reader review â stress tests â verification, with a dual-persona read for the weak-reader set.
written for weak readers
Front-loaded prohibitions, positively-phrased instructions, and VERIFY checks a correct run can actually satisfy.
gaps stated out loud
Desk-checked material is labeled, hardware numbers are hedged, and uncovered topics are listed rather than implied.
a design record
A plan document captures every scope and boundary decision, so the library stays coherent as it grows.
under the hood
- router-description contract â each skill's description says when to trigger, the literal keywords, and when NOT to, in â¤500 characters.
- rigid templates â fixed section structures per audience; the rigidity is a feature weak readers depend on.
- real-model verification â weak-reader skills were stress-tested against a live 7.5B model and a local embedder.
- canonical wordings â safety-critical phrasings are owned by one skill and repeated verbatim, never paraphrased.
- failure-mode catalog â the recurring defect classes review kept finding are written down so edits don't reintroduce them.