<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
tjbai/dotfiles/style
──Details
PublishedOctober 21, 2026 at 09:41 AM
Content Hashsha256:c9aaddf971bc4229...
Git SHA3cbc0f556251
──Files
Files (1 file, 2.7 KB)
SKILL.md2.7 KBactive
SKILL.md · 43 lines · 2.7 KB

version: "1.0.0" name: style description: "The global style index for everything written: code, Python, commits, PR titles and descriptions, and technical prose. Sets the default register and the de-slop bar. Apply on the first pass, not as a cleanup pass. Use whenever writing or cleaning code, committing, writing PRs, or writing any prose. Triggers on: style, slop, de-slop, deslop, clean up, tidy, commit, PR, prose, my style."


Style

Index skill. Before producing an artifact, read the matching reference file. Apply it on the first pass — a "de-slop" or "clean this up" request from the user means this skill failed.

ProducingRead
Any codereference/coding.md
Pythonreference/python.md, after coding.md
Git commitsreference/commits.md
PRs: titles and descriptionsreference/pr.md
Prose: docs, ADRs, scratch notes, chat, commentsreference/writing.md

When the user says "de-slop" or "clean up", apply coding.md and python.md.

Always

  • The default prose register is the writing-simple skill (STE sentences, closed NGSL vocabulary). Load it for any durable prose. Details and checker policy: reference/writing.md.
  • Never hard-wrap prose at a column width. Write each paragraph as one line; let the renderer soft-wrap. When editing hard-wrapped text you touch, unwrap it. Exceptions: code, tables, YAML/TOML/JSON, formats that genuinely require wrapping.
  • Zero inline comments in code by default. See coding.md for the rare exceptions.
  • Smallest accurate diff. Never reformat lines you did not change.

Amending this skill

When the user nudges on style during a session, this skill is stale or wrong. Do not just comply and move on.

Trigger. Two or more corrections on the same theme in one session, or one correction that contradicts a rule written here.

Action. At the next natural pause (end of the task, before the final report):

  1. Propose a concrete amendment as a diff to the right reference file. Quote the user's correction verbatim as evidence, typos included.
  2. For a theme no reference file covers, propose a new reference/<topic>.md plus an index row above.
  3. Keep it minimal: one rule per correction pattern. No restructuring, no rewriting adjacent rules.
  4. Apply only after the user approves.
  5. ~/.config/agents/skills is the source of truth but not a git repo. After editing, run ./update in ~/dev/dotfiles and commit there with a lowercase subject: style: <rule added or changed>.

Never propose an amendment for a one-off, context-specific instruction. Amend only for corrections that would apply to future sessions.

All versions