Founding TenHalf price website?it's a steal…
Journal

Decision 0008

Which design skills build the journal pages

In plain English

Before an agent designs the journal pages on method7.co.uk, it needs guidance on how to design well, and on what this particular site looks like. We chose one design skill from Anthropic for the first job and wrote our own brief for the second. Nothing new is installed.

Context

A skill is a short instruction sheet an AI agent loads when a job calls for it (decision 0005). The journal has to look considered rather than templated, but it lives inside an existing site with its own colours, fonts and automated checks, so it must not become a different brand. Third-party skills are code written by strangers, so each candidate was read in full before being judged, and none was installed.

Options

  1. frontend-design, from Anthropic (github.com/anthropics/skills, Apache 2.0 licence, already installed). One instruction file and no scripts. It teaches a way of working: ground the design in the subject, plan colours, type and layout before writing code, check the plan against the brief, spend boldness on one element, and critique the result. It lists the habits that make pages look machine-made. Two catches: its own text uses US spelling and nine em-dashes, and one of the “generic” looks it warns against, near-black with a bright acid green, is Method7’s actual brand. It also says a brief that sets a direction must be followed exactly, which resolves that.
  2. brand-guidelines, from Anthropic (same repository, Apache 2.0). No scripts. It applies Anthropic’s own colours and fonts, mainly to slide decks. Wrong brand, so unusable as it stands, but its shape (one file stating the palette, the fonts and when to use each) is a good pattern for a Method7 version later.
  3. webapp-testing, from Anthropic (same repository, Apache 2.0). Ships Python scripts that start a web server and drive a browser to take screenshots. It tells the agent to run its scripts without reading them first, which breaks our rule for outside code. It also needs a browser library the site does not have, and the browser tools Claude already has can take the same screenshots.

Two more in that repository were ruled out on their descriptions alone: theme-factory applies preset colour themes, and web-artifacts-builder runs scripts that install a different set of web tools.

Decision

Option 1, for process only, with uk-house-style for every word on the pages. What to build comes from journal/design-brief.md, which is written against the site’s own design tokens, components and checks, and overrides the skill wherever they disagree. We decline options 2 and 3.

We are not writing a Method7 design skill yet. The brief covers one build. If a second page build needs the same guidance, the brief’s reusable parts (tokens, components, what the site’s checks demand) become a skill, shaped like option 2.

Consequences

  • Nothing to install or vet beyond what is already here.
  • The builder has two sources of design guidance and must let the brief win. The brief says so at the top.
  • The skill’s US spelling and em-dashes could leak into copy. The house-style skill and the em-dash check on both repositories catch that.
  • Writing the brief found a fault in what is already published: decision 0004 links to a private research file, which the site’s link check would reject. The brief tells the builder to handle it; the entry itself needs a separate fix.
  • Accepted by Simon on 12 November 2026.

Thinking about agents in your own business?

Tell me which jobs eat your week. I'll tell you plainly whether AI agents would help, and what building it would involve.