Compact readme

compact-readme

Take an existing README that has grown into a wall of text and tables, and tighten the landing view so a reader reaches the install/run command in ~30 seconds — while preserving every fact by relocating (never deleting) detail into <details> blocks and merged paragraphs.

[!IMPORTANT] This is the complement to write-readme. write-readme = scaffold a new README from scratch in the kirchDev house style. compact-readme = slim down an existing one. If there is no README yet, stop and use write-readme instead.

[!IMPORTANT] Never delete content. Everything you cut from the visible view must land somewhere — folded into a <details> block or merged into an existing paragraph. The goal is fewer visible lines, not less information.

Goal & rule of thumb

• The un-collapsed (visible, outside <details>) line count should roughly halve, with zero information loss.
• Reference detail stays in the file, just collapsed or consolidated.
• Reference: the envprism README went from ~194 → ~98 visible lines by folding the feature list + config tables into <details>, merging Install+Quickstart, and deduping the tagline / runtime note / "round-trip preserving" claim. See REFERENCE.md for the worked before/after.

Step 1 — Diagnose (measure before you cut)

1. Read the whole README and count total lines (wc -l README.md).
2. Attribute the bulk — which sections eat the most lines? Usual culprits: long feature bullet walls and big Configuration/Options tables.
3. Spot redundancy, e.g.:
   • the tagline restated in hero + hook + a later sentence,
   • a runtime/requirement note ("needs Bun, npx won't work") repeated across hero, install, and a callout,
   • a repeated core claim ("round-trip preserving", "zero-config") stated in several places,
   • a block that mirrors a file the repo ships — every config option out of a schema, every script out of package.json, every env var out of .env.example. This one is not merely long: it is duplicated state, so it is already wrong somewhere or will be at the next commit.
4. Check the house style is intact (centered hero, section emojis, Versioning/License closers) — preserve whatever is already correct.

Present the diagnosis (line count, biggest sections, redundancies found) and the planned cuts before editing — see Interaction.

Step 2 — Transform

Apply these, in roughly this priority:

1. Feature list → ~8–10 real highlights + a <details> superset. Trim the visible list to genuine highlights. Move implementation detail (regexes, internal parser rules, full theme/color-key lists, every single keybinding) out of the highlights. Put the complete, categorized list into <details><summary>Full feature list</summary>…</details> as a superset — nothing is lost.
2. Configuration/options → teaser + minimal example + <details>. Keep a short teaser sentence and one minimal config example visible; collapse all option tables and theme/color keys into <details><summary>All configuration options</summary>…</details>. Keep everything in one file — do not spin out to docs/ unless the user explicitly asks.
3. Merge Install + Quick start into one "Install & run" section. State the runtime requirement exactly once, as a > [!IMPORTANT] callout in that block, including the concrete gotcha (e.g. "npx won't work").
4. Dedupe repeated claims — pick the single best spot for each restated tagline / requirement / core claim and remove the echoes.
5. A block mirroring a file → the values that matter + a link. Where a table was transcribed from a schema, package.json or .env.example, <details> is the wrong tool: collapsing a copy keeps its drift and merely hides it. Keep the two or three values a newcomer actually sets, then point at the file that owns the rest. Nothing is lost — the content was never the README's to hold; it is going back to the one place that cannot go stale. Name the file in the plan so the move is visible.
6. Reduce bold density. If every bullet is fully **bold**, nothing stands out. Bold only ~the first phrase of each bullet.
7. Tighten prose — e.g. a Why/Problem section from 3 paragraphs to 2 — but keep the author's voice; do not flatten it into generic AI slop.

Step 3 — Target structure (order)

Hero (logo, tagline, badges, optional demo GIF)
Hook (1 code line + ~2 sentences)
🤔 Why (problem → solution, tight)
📦 Install & run (merged; requirement callout ×1)
✨ Features (~10 highlights + <details> full list)
[optional] 🧪 CI / scripting (short standalone block, if it deserves one)
⚙️ Configuration (teaser + example + <details> tables)
🤝 Contributing
🛣️ Versioning
📄 License

Use the prescribed section emojis from write-readme's REFERENCE (Section emoji catalogue).

<skills-plan>

Presenting the plan

Everything this skill puts in front of a human — plan, preview, candidate list, findings report — is read once, in a terminal, and answered there. So every section of it renders on arrival, with no interaction needed to reveal it: prose, lists, tables, fenced code.

Never fold content behind a control. <details>/<summary> is a browser widget, and a terminal has no way to open it: the summary line prints and everything under it does not. The plan then arrives as headings with nothing beneath them, and the failure is silent on both sides — the skill believes it reported, and the reader sees no marker saying anything is missing, so a human confirms a plan whose contents never reached them. What gets folded is whatever ran long, which is to say the part the decision actually rested on. The same holds for anything else needing a click: a tab strip, an accordion, a "show more".

Length is handled by shortening, never by hiding. This is a fixed rule of the skill, not a per-run judgement, so it holds however long the list runs. Trim to what the decision needs, group the rest by something the reader already thinks in (ecosystem, kind, verdict) with a count per group, or split it across sections. What is left out is left out visibly: say how many, why, and the exact command that shows the rest.

This binds what the skill presents, not what it writes. A <details> block inside a README, an issue body, a pull request description or a docs page is rendered by a browser and is entirely legitimate there. The rule is about the message a human reads to decide — never about the content of a file.

</skills-plan>

Guardrails

• Never delete content — relocate into <details>, merge into a paragraph, or (only for a block copied out of a file the repo ships) replace it with a link to that file. Content leaves the README only when something else in the repo already holds it.
• Preserve the house style strictly — centered hero in <div align="center">, section emojis, Versioning/License at the end. For kirchDev repos, treat write-readme as the style source of truth.
• Leave these untouched: badges, links, the demo GIF, and the license/author line.
• Factual accuracy — never rewrite or invent defaults, flags, or option values. Copy tables verbatim; only move them.
• Formatter check — if the repo has an oxfmt/Prettier config, confirm README.md is excluded from the formatter (kirchDev excludes it in lint-staged.config.js, because oxfmt reflows the centered hero and <details> markup). If it is not excluded, run the formatter after editing so you don't leave a diff CI will reject.
• Report visible-line delta after the rebuild: before vs after (outside <details>), plus a short "what moved where" list.

When not to / when to ease off

• Don't create new READMEs — that's write-readme.
• If the README is already lean, don't force structural surgery. Just smooth redundancy and bold density, and say so in the report.

Interaction

Before writing, show the diagnosis + planned cuts, then ask:

1. Aggressiveness — light / medium / aggressive.
2. Full feature list — collapse into <details> (recommended default) / drop entirely / move to docs/ (only offer docs/ when the repo isn't docs-disabled — skip it when the resolved config sets docs: false; REFERENCE.md states how to resolve it).

Then apply, and end with the before/after visible-line count and the "what moved where" summary.

Reference

• Transformation playbook, <details> patterns, and the envprism before/after: REFERENCE.md.
• House-style source of truth (emojis, badges, closers): write-readme and its REFERENCE.md.