Press

Codex runtime

When running in Codex, invoke this skill as $press. Resolve scripts, assets, and references from the directory containing this SKILL.md, regardless of the current working directory. Existing ~/.claude/ personal-data paths remain valid and are still used by the bundled scripts; they do not require Claude to run. Map Read/Write/Edit/Bash to the available file and shell tools, and WebSearch/WebFetch to available web tools. For AskUserQuestion, use an available question tool or a concise chat question; wait for answers that gate action. Use Codex's delegation tools for required subagents when available; otherwise disclose that independent execution is unavailable. Discover connected apps by capability rather than assuming Claude MCP tool names exist.

/press — the brand system

You are running the press skill. PRESS is the warm-paper editorial brand every artifact shares: a morning brief, a city profile, a budget report, a résumé, a LinkedIn card, a dev-log cover, and the site all read as issues of one publication.

Announce at start: "I'm using the press skill for the brand."

All commands below run from the directory containing this SKILL.md (the skill's install dir, $SKILL_DIR). Resolve it once and cd there first. When working in another repo, pass --repo <path>.

The one rule

Never write a brand value into a file by hand. Not a hex, not a font stack, not the monogram. Every consumer gets its values generated into a marked region, and CI fails if any of them drifts. Copying a value is how this brand ended up in eight hand-ported copies with five different names for the same orange, which is the problem this skill exists to end.

If a consumer needs a color that doesn't exist yet, add it to brand/tokens.json and re-emit. Never invent one locally.

What's here

Composing something new

1. Read brand/laws.md first, then brand/components.md. They are short. Composing without them produces something that uses the right colors and still looks like a different product.
2. Pick the components the content needs, not a template. PRESS is a brand system, not a template — two artifacts should never share a skeleton just because they share a brand.
3. Get the values from the CLI, never from memory:

   node bin/press.js tokens --format css     # a :root block
   node bin/press.js tokens --format json    # raw values
   node bin/press.js tokens --format md      # the palette as prose
   

4. Lint before showing the user:

   node bin/press.js lint <file…> [--accent-cap 2]
   

   Findings are mechanical brand violations, not style opinions. Fix them.
5. Show the rendered artifact, not a description of it. Per brand/agent-ui.md, a visual claim without the image in the transcript is not a result.

Checking every consumer is in sync

node bin/press.js check                 # this repo
node bin/press.js check --repo ../budget
node bin/press.js doctor                # the whole registry, present or not

check fails on three things, not one: a region whose bytes drifted, a region that has gone missing from a file that should have one, and a run that resolved zero targets. The last two matter most — a checker that verifies nothing reports "all clean", which is exactly how a gate turns decorative.

On failure it prints the diff and the exact press emit command that fixes it. Run that; do not hand-edit the region.

How a change reaches every product

There are two different questions, and conflating them is what leaves a consumer silently stale.

A pinned check can never answer the second question — it passes forever against the version it was pinned to. That is not a hypothetical: this skill's own site consumer sat two releases behind with entirely green CI.

Consumers pin on purpose. A mutable @latest in a repo that auto-deploys to production is a supply-chain hole, and a reproducible check is the only kind worth gating a merge on. So freshness is pushed from here rather than polled by each consumer:

node bin/press.js propagate --repo ../budget --dry-run   # is it behind?
node bin/press.js propagate --repo ../budget             # re-emit + bump its pin

propagate re-emits every region it resolves, bumps any @natjswenson/press@<version> pin it finds in that repo's workflows, and reports what moved. The caller turns that into a pull request, so nothing lands unreviewed — but nobody has to remember.

One version per repo. A consumer's CI pin, its region receipt and the current release are kept equal, so "is this repo up to date?" is answerable by looking. The earlier rule — only act when values move — left three divergent version numbers per repo and no way to tell a healthy consumer from a stale one.

Every release therefore opens a PR in every consumer. What keeps that readable is that the PR says which kind it is:

.github/workflows/press-propagate.yml runs this on every release and weekly. It propagates to every long-lived branch a consumer has — dev and the default branch — because landing on dev alone leaves the branch that actually ships stale until someone remembers to promote, and makes the next promotion carry an unrelated brand diff.

Each consumer's README carries the release in a generated region — press:masthead for a skill, press:version for a product repo — so the version a repo is on is visible without opening a workflow file.

Changing a brand value

This is the one operation that touches every product, so do it deliberately.

1. Edit brand/tokens.json. One value, with a reason.
2. node bin/press.js emit --dry-run and read what would change.
3. node bin/press.js emit, then run the affected skills' own test suites — a token change must not break anyone's baseline.
4. Re-render one real artifact per affected medium and look at it.
5. Bump the version and add a CHANGELOG.md entry in the same change.

Repos outside this one need no manual step. Releasing the version bump fires press-propagate.yml, which opens a PR in every consumer whose bytes actually moved. To see it early, dry-run against a local checkout:

node bin/press.js propagate --repo ../budget --dry-run

Onboarding a new consumer

1. Add a target to targets.json: id, repo, path, region, syntax (python | yaml | css | md), emitter, and params.

2. Choose the emitter:

   Emitter	Produces
   python-theme	the token dict plus the shared deep-merge loader
   python-consts	flat module constants, for a script with no override file
   css-vars	a custom-property block, with per-consumer aliases
   md-palette	the palette as a prose bullet list
   markdown-block	one of the brand docs, inlined into a SKILL.md
   version-badge	a "this repo is on press vX" note for a non-skill README
   readme-masthead	the masthead a skill README wears
   gha-header	the masthead a generated GitHub Actions workflow wears
   json	raw values

   gha-header comments its own body. Every other emitter returns code in the target's language and lets renderRegion comment just the two markers; a workflow masthead is commentary throughout, so emitting it bare would splice box-drawing into the document and produce YAML that cannot parse.

   readme-masthead is for a skill README; version-badge is for every other README. The masthead carries the full .mast — stamp, brand line, document kind, issue, byline — and the rule under it, and it requires the README's H1 to be the bare skill name so its ^# <name>$ anchor is the same everywhere. A product repo's README has its own title and gets the badge instead.

3. Pick the font profile for its rendering engine. params.font_profile is browser by default. Anything going through fontconfig — WeasyPrint above all — must say fontconfig, because it walks the fallback chain for real and the browser chain resolves headlines to Helvetica Neue Heavy Condensed. Survey the consumer's current stacks before you assume.

4. Add an init anchor naming the first and last line of the hand-written block the region takes over, so the duplicate is swallowed, not left behind.

5. node bin/press.js emit --target <id> --init

6. Add a press check step to that repo's CI, pinned to an exact version: npx -y @natjswenson/press@0.5.0 check --repo .. Never write a bare npx @natjswenson/press — with no version reference at all, npx silently prefers a stale global install over the registry, which cost this repo a release once already. An exact pin also keeps a mutable dependency out of a repo that auto-deploys to production; propagate bumps it for you.

7. If its GitHub repo name differs from the repo field (budget lives at local-budget), add "github": "<name>" so propagation can find it.

Survey before you emit. Diff the consumer's real values against the token set first: adopting the site turned up a 'JetBrains Mono' fallback and three alpha tints the tokens could not express, and emitting blindly would have silently dropped them. Widen the tokens to the union; never narrow a consumer.

Aliases are deliberate. The résumé keeps --sig, the site keeps --fg. The names stay idiomatic to their medium; only the values are shared. Renaming across three repos would be churn for no gain.

Never generate a whole file. The region owns the token block; the consumer owns everything else — its 250-line stylesheet, its poster geometry, its personal avatar footer. A whole-file sync clobbered exactly that once already.

Adopting the prose contracts

brand/agent-ui.md is the single runtime presentation contract. Skills carry only a stable dependency note telling the host to load the installed PRESS skill before producing user-facing output; they do not embed, version, or copy the contract. This keeps UI changes in one file. brand/voice-core.md remains a content rule for artifacts and is still consumed where a skill explicitly needs it.

There is no native cross-skill import field in Claude or Codex manifests. The dependency note is therefore the host-neutral handoff: Claude Code loads /press, Codex loads $press, and the host reads brand/agent-ui.md before rendering the run.

A medium's own voice layers on top and wins on conflict — ghostwriter's learned profile and devlog's release-note shape stay where they are. The core is what applies when nothing more specific does.

What press does not decide

Layout and composition (the medium's business), chart color validation (the dataviz skill), and any voice learned from the user's own writing. PRESS sets the floor everything shares, not the whole of anything.