Skills

Agent skills for real engineering work.

These skills are small, easy to adapt, and composable. They work with any model. Hack around with them, make them your own.

Skill	What it does
eliminate-no-op	Audits an instruction file for directives that burn context without changing behavior, and proposes a rewrite beside the original
evaluate-skill	Builds an eval suite for another skill, runs it, reports — optionally with a loop that revises the skill from its own failures

The second one measures the first: evals/eliminate-no-op/ is a real suite with a measured baseline, and it is where both skills were shaken out.

Install

With the skills.sh installer:

npx skills@latest add pat-lewczuk/skills

Or copy the directory you want straight into your project:

cp -r skills/eliminate-no-op /path/to/your-project/.claude/skills/

Scripts are stdlib-only Python 3 — there is no install step, and nothing to keep in sync.

Skills

eliminate-no-op

Audits agent instruction files — AGENTS.md, CLAUDE.md, SKILL.md, .cursorrules, system prompts — for directives that burn context without changing behavior.

Instruction files rot. Every incident adds a rule, no one ever removes one, and the file drifts into a wall of well-meant text the model already agreed with. That text occupies context on every turn and dilutes attention away from the rules that actually encode project knowledge.

The skill runs a mechanical pass (duplicate detection, dead path references, token cost per section), then classifies every directive as KEEP / TRIM / MERGE / MOVE / CUT against five survival tests. It produces a findings report and writes the rewrite beside the original — never in place — so you review a diff before anything is applied.

It is deliberately conservative: deleting a line that was quietly preventing a recurring bug costs far more than leaving three redundant lines alone.

What it recovers

Measured on the eval suite's seven audit cases, from the rewrites the skill actually produced (baseline run, 2026-07-26):

Case	Target	Lines	Est. tokens	Saved	Cut
bloated-webapp	AGENTS.md	129 → 63	1,229 → 527	702	57.1%
skill-md-target	SKILL.md	94 → 61	930 → 492	438	47.1%
duplicated-hierarchy	AGENTS.md	59 → 33	436 → 249	187	42.9%
no-repo-paste	colleague-AGENTS.md	48 → 36	471 → 329	142	30.1%
stale-migration	.cursorrules	19 → 12	316 → 194	122	38.6%
long-but-dense	AGENTS.md	93 → 88	1,271 → 1,172	99	7.8%
lean-cli	AGENTS.md	40 → 37	424 → 406	18	4.2%
Total	7 files	482 → 330	5,077 → 3,369	1,708	33.6%

Read the spread, not the mean. The bottom two rows are the healthy files, whose labels expect 0–25% and 0–22% — a large cut there scores as a failure. Maximising tokens removed across this table would mean destroying them, which is why the suite scores cut size against a per-file band instead of rewarding volume.

What a saving is worth depends on where it lands. AGENTS.md, CLAUDE.md and .cursorrules load on every turn, so 702 tokens off bloated-webapp is 702 tokens back on every turn in that repo, permanently. A SKILL.md body only loads when the skill fires — there the resident cost is the description field, which that audit cut from 152 tokens to a tight trigger sentence.

Two things the table undercounts: duplicated-hierarchy also recommended deleting CLAUDE.md outright (~135 tokens, four of its six lines duplicated the root file), and directives verdicted MOVE leave the always-loaded root for a nested file that loads only in its own subtree.

Figures are the len/4 estimate analyze.py uses. Files this dense with backticked paths and flags tokenize worse than plain prose, so treat the absolutes as ±15%; the ratios hold. Each audit cost $0.73–$1.04 and 13–18 turns to produce.

evaluate-skill

Builds an eval suite for another skill, runs it, and reports — optionally with a loop that revises the skill from its own failures.

A skill is a prompt, and a prompt is untested code. The usual way to improve one is to read it, decide a paragraph feels weak, and rewrite it. That has no error signal: the edit that makes a skill read better and the edit that makes it work better are not the same edit.

Given a skill name it asks three things — what the skill must get right and what it must never do, whether you want to supply the cases or have them generated, and whether you want the improvement loop. Cases you describe in the chat get converted into real ones, copying the workspace from a real project if you point at it; a failure you actually remember is better ground truth than anything a model invents.

From there it scaffolds the suite, verifies the instrument against a do-nothing run and a perfect run before spending anything, runs it, and writes an HTML dashboard. Ask for the loop and it additionally revises the skill, rejects patches that memorised the fixtures, and keeps a revision only if it holds up on cases it was never tuned against.

It refuses to start without a named target skill.

python3 scripts/scaffold.py --skill my-skill --case happy:train --case restraint:train \
  --case unseen:holdout --case triggers:train:trigger

The suite it builds scores checks and rubric criteria into weighted metrics, with safety heaviest — for nearly every skill the expensive failure is confident damage, not a missed improvement — and gates that zero a case outright. selftest.py is what makes the numbers worth reading: it scores a run that produced nothing (must be near zero) against one built to satisfy every check (must be near one), so a case that cannot separate those two gets caught before you spend anything on it.

Evals

evals/eliminate-no-op/ holds the eval suite for that skill: eight labelled cases, weighted metrics that punish deleting a load-bearing line three times harder than missing a platitude, and a loop that revises the skill from its own failures and keeps a revision only if it holds up on cases it was never tuned against.

cd evals/eliminate-no-op
python3 harness/selftest.py     # validate the graders, free
python3 harness/run.py          # run the suite
python3 harness/loop.py --iterations 2
python3 harness/report_html.py  # regenerate results.html + the image above

Details, case-by-case, in evals/eliminate-no-op/README.md.

License

MIT