skills: context-frugal skill discipline, and a lint that counts what a skill costs

[rookslog]

Motivation

A local log study (2026-08-13, unpromoted — operator transcript corpus, 3,887 files, ~92k requests) found that user-facing prose is roughly 3% of list-priced spend while context volume (cache reads and writes) is roughly 75%. The single worst observed event was a bundled skill that injected ~70k tokens of inlined reference to answer a four-number lookup. Skill delivery cost, not output verbosity, is the dominant style-adjacent cost lever.

Proposal

Progressive disclosure is a principle of style. Ship it as one:

1. A craft-tier skill teaching context-frugal skill authoring:
   • The frontmatter description is trigger text only. It rides in every session, so every word there is a standing tax.
   • The SKILL.md body carries triggers and procedure under a stated token budget. Everything else moves to references/, read on demand.
   • A body never inlines what a file on disk can answer with a targeted read.
2. A lint check over skill files: estimate SKILL.md and description token counts, warn past a budget. Warn, never error — a large skill is sometimes right, and the entry must state why (same three conditions as word lists).

Measurement contract (per #96)

• Construct: delivery cost of a skill without loss of trigger accuracy or task performance.
• Measures: resident description tokens, body tokens, per-invocation context cost.
• Counterweight: trigger accuracy and task success (bench arms). Cutting tokens that costs triggering is a regression, not a win.
• Falsifier: a promoted study showing budgeted skills trigger or perform worse than unbudgeted ones.

Grounding: the skill gets a matrix in grounding/craft/ like any other. The lint gets tests and joins npm run check as a named script.

[rookslog]

Probe design: citation-only vs full skill (the top rung of the frugality ladder)

Question. Does citing a standard buy the compliance that shipping the skill body buys, and on which models. Matt Pocock's wait-what delivers STE in three lines by citing ASD-STE100 and pointing at a project glossary. If that holds on large models, this issue's top rung becomes "cite resident, load the body on demand". If it fails only on the dictionary, the dictionary becomes the on-demand module and the prose rules stay citable.

Arms.

• C0 — no guidance. The control from the probe protocol.
• C1 — bare citation. A three-line skill whose body says: write this per ASD-STE100.
• C2 — the shipped simplified-technical-english skill, unchanged.

Same writing tasks across arms. Two model blocks, one opus-class and one haiku-class, because the conjecture is that the citation-only gap widens as the model shrinks. Each arm runs under --setting-sources user with a throwaway home, per bench/probes/README.md.

Scoring, mechanical first.

• Dictionary compliance: word-pair violations per 100 words. This is the part of the standard a citation is least likely to recover, because the approved pairs are the licensed and least-memorized part. Depends on Use the new quotation permission: quoted rules in the matrix, then a word-pair lint dictionary #19. Until Use the new quotation permission: quoted rules in the matrix, then a word-pair lint dictionary #19 ships, a person scores this dimension and the reading is recorded per the editorial discipline.
• Lintable prose rules: sentence length, one instruction per sentence, the word list. stylewright lint scores these today.
• A person reads a sample from each arm. No arm grades itself.

Contract, per #96.

• Construct: the compliance delta between C1 and C2, per model block.
• Counterweight: delivery cost. C1 is roughly a hundred resident tokens. C2 is a multi-thousand-token load, measured by the usage instrument's census join rather than assumed.
• Decision rule, stated before collection: if C1 holds within a stated margin of C2 on the opus-class block and the dictionary dimension does not collapse, the craft skill's top rung is cite-then-load. If the dictionary collapses under C1 on every block, the dictionary is the on-demand module and the rule still holds for prose.

Pre-registered priors, all conjecture. C1 recovers the flavor of the standard — short sentences, active voice — because the identifier is a strong pretrained pointer. C1 misses dictionary-level word choices. The gap widens on the small block and in long contexts. A long-context arm is out of scope for the first pass.

(Draft by the session agent, 2026-08-13. Nothing here authorizes collection — the measurement design governs promotion as usual.)