chore(skills): enforce soft-wrap in every md-writing skill & agent

[skapoor8]

Intent

Make every skill and agent that emits prose markdown state the soft-wrap rule (one line per paragraph, no mid-paragraph hard-wrap "endlines"), so generated .md artifacts reflow correctly and produce clean diffs. Several skills already had it (spec, dev, em, pm, ux, docs, ed, qa); coverage was inconsistent and most agents — the code that actually writes the files — lacked it. This closes those gaps.

The problem (rationale for review)

LLM-generated markdown sometimes hard-wraps prose (manual newlines mid-paragraph at ~80 cols). That:

• reflows badly on narrow screens and non-reflowing renderers (a paragraph becomes a ragged column), and
• corrupts diffs — a one-word edit re-wraps the whole paragraph, so the diff shows every line changed.

The repo convention is soft-wrap: each paragraph is one continuous line; the renderer wraps. Newlines still separate paragraphs, list items, headings, and code fences.

Why embedded instructions, not a wired-in tool

You asked whether a "standard tool" is the correct way. Key consideration: these skills are installed into other people's repos (npx skills add …) and run there. A formatter config or mise task in this repo cannot enforce anything in the repo where a skill actually executes — only the instruction embedded in the skill/agent travels with it. So the portable, always-on mechanism is the embedded rule, and that's what this PR makes ubiquitous and identically worded (no drift).

The deterministic sanitizer still has a place: the standard tool for "collapse hard-wrapped prose to one line, preserving code fences / lists / tables / front-matter" is Prettier with prose-wrap: never (prettier --prose-wrap never). A naive sed/fmt line-join would corrupt code blocks and tables, so Prettier is the right one. Each added rule now names it as the optional deterministic pass a consumer can run in their own repo/CI — but we deliberately do not hard-wire it as a per-skill runtime dependency (it would not be reliably present at skill-run time and would add latency to fast skills).

Changes

• Skills (added the Critical Rule where missing): pr, flow, skill, usage, gh, glab, changelog, linear.
• Agents (added a "Markdown output" note — these are the actual writers): spec/* (3), pm/* (7), pr/* (3), skill/* (3), dev/{researcher, linear-tasks-agent}, em/linear-push-agent — 19 agents.
• Wording reused verbatim from spec/SKILL.md so the rule is identical everywhere.

Deliberately excluded (please sanity-check at review)

• git commit messages — commit bodies follow the ~72-column wrap convention, so soft-wrap would be wrong there. The git skill is untouched.
• Code-pattern / context / lookup skills (aws, cpp, docker, gcp, hpc, llm, mise, mle, python, react, sveltekit, tanstack, terraform, typescript, compgeo, cz, icons, vim, odin, task/tasks) — they help write code or are keybinding lookups; they do not emit prose markdown artifacts.

Validation

• mise run skills:validate — all pass
• Coverage verified: 8 skills + 19 agents now carry the rule; git confirmed untouched

Future work

• If you want CI-level enforcement on this repo's own committed markdown, a follow-up can add a .prettierrc (proseWrap: never) + a mise fmt:md task — kept out of this PR because it is a separate, opt-in concern from the portable per-skill instruction.

[skapoor8]

Addressed the one review finding: added the soft-wrap rule to the deprecated dev-docs alias for parity (it still ships a doc-writing workflow). Review confirmed all 19 agent appends land at clean EOF, the 8 skill insertions are valid markdown, coverage is complete, and git + code/lookup skills are correctly untouched.

[github-actions]

🎉 This PR is included in version 1.67.1 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀