Skip to main content

The content-architecture skill

An audit of your site's copy and content structure, followed by the restructuring. It hunts the problems that accumulate as content grows - the same feature described three different ways, whole sections doing another page's job, journey steps with no surface, pages with three CTAs or none - presents them as a visual HTML report, and then fixes whichever finding you pick.

The goal it works toward: every message has exactly one home, every page has one job, and the reader's journey has no dead ends.

Invoke it with /content-architecture (Claude Code, Cursor) or $content-architecture (Codex).

Install​

The skills package is the default path:

npx @reactor-tools/skills add content-architecture

Or run npx @reactor-tools/skills with no arguments for the interactive picker.

Prefer the raw files? This skill is two documents - the skill and its report scaffold:

mkdir -p .claude/skills/content-architecture
curl -o .claude/skills/content-architecture/SKILL.md https://docs.reactor.tools/content-architecture/SKILL.md
curl -o .claude/skills/content-architecture/HTML-REPORT.md https://docs.reactor.tools/content-architecture/HTML-REPORT.md

What it looks for​

Seven finding types, each named for what it costs the reader: drift (one capability, three phrasings), squatting (a message with no home, restated ad hoc everywhere), redundancy (a section doing another page's job), journey gaps (awareness → evaluation → adoption steps with nothing serving them), CTA conflicts, voice fractures (register changing mid-site against VOICE.md), and orphans (pages nothing links to). Findings that are just stylistic preference don't make the cut - each must name a reader-visible cost.

How a session runs​

  1. Map. The agent walks every prose surface in the repo (marketing pages, docs, README, UI strings, emails - plus any hosted URLs you give it) and builds the message map: every claim, everywhere it appears, phrased how.
  2. Report. Findings render as a self-contained HTML file in your temp directory and open in your browser: a site map colored by finding, one card per candidate fix with side-by-side evidence and a before/after visual, each badged Strong, Worth exploring, or Speculative, ending with the top recommendation.
  3. Grill. You pick a candidate; the agent interviews you one question at a time (which surface becomes the message's home, what the other sites say instead, what redirects follow) until the restructuring is fully decided.
  4. Restructure. The agreed change lands in your files, and the message map re-runs over every touched surface to confirm the finding is actually gone.

This is the manual version. Reactor runs the same discipline automatically against your saved brand voice, Guard rules, and performance data.