Betterskills
Generated for Codex from the canonical source plugin. Run
just updateat the marketplace root to refresh it.
Critique, improve, and create Codex skills against one standard: a skill is a correction you got tired of making. Target the step that fails, use leading words that recruit training priors instead of coined terms, keep skills small and user-invoked, grow them only from observed corrections, and delete every instruction that does no work. critique reports findings with fixes and edits nothing; improve prunes no-ops, swaps explanation for leading words, and folds recurring corrections from session history in as one-sentence fixes; create mines session history for repeated corrections, checks overlap and default behavior before writing, and drafts a minimal skill. Subcommands: critique, improve, create.
Install with Codex
Register the repository marketplace with codex plugin marketplace add ., then run codex plugin add betterskills@florianbuetow-codex-plugins or install it from Florian Buetow Codex Plugins in the Plugins Directory. Invoke its workflows with $<skill-name>.
Critique, improve, and create Codex skills against one standard: a skill is a correction you got tired of making.
Usage
$betterskills critique grill-me
$betterskills improve resume:review
$betterskills create "it keeps writing the code before we agree on the plan"
$betterskills:critique all
$betterskills:improve my-skill precise
$betterskills:create
Skills
| Skill | Purpose |
|---|---|
betterskills | Router — sends the request to one subskill |
betterskills:critique | Judges a skill (or all of them) against the standard; reports findings with fixes, edits nothing |
betterskills:improve | Rewrites a skill: prunes no-ops, swaps explanation for leading words, folds observed corrections in as one-sentence fixes |
betterskills:create | Mines session history for repeated corrections, checks overlap and default behavior, drafts a minimal user-invoked skill |
The standard
references/standard.md holds five rules, each with checks (R1.a … R5.c)
that every finding cites:
- Target a correction you keep sending back — aim at the step that fails; guard against asset rush.
- Use leading words — established terms recruit training priors; coined terms cost definition tokens.
- Keep it small; invoke it deliberately — prefer
disable-model-invocation: true. - Grow it from real use — every instruction traces to an observed failure; revise weekly.
- Delete instructions that do no work — no-ops, prohibitions that could be stated positively, overlapping skills.
references/corrections.jq lists the corrections you made in
~/.claude/projects session logs, so create and improve work from
evidence rather than guesses. It needs jq.