PageKit

Skills for making web pages that are contextually relevant to your audience and what your product does.

PageKit ships as a Claude Code plugin. One install, fourteen skills (twelve method skills + two subagents), and the full 7-step chain runs end-to-end: signal doc → message spine → first-page decision → page argument shape → proof map → first-page draft → claim check. An anti-slop regression script catches mechanical slop patterns; the claim-check subagent catches the rest.

If you do not use Claude Code, there is a chat-paste path in prompt-paste/.

Install

Inside Claude Code:

/plugin marketplace add hnshah/pagekit
/plugin install pagekit@pagekit

That is the whole installation. All fourteen skills register automatically.

Use it

Ask Claude Code to run PageKit on any object:

Run PageKit on <product or idea>.

The orchestrator skill (pagekit) walks you through the seven steps, scaffolds a fully-logged runs/<name>/ folder, and invokes each step skill in turn.

To run one step on its own, invoke the step skill directly:

Use pagekit-signal-doc on <product>.

The skills

Skill	What it does
pagekit	Orchestrator. Runs all 7 steps end-to-end into a fully-logged run.
pagekit-signal-doc	Step 01. Turn source briefs into a signal doc.
pagekit-message-spine	Step 02. Reduce the signal doc to a message spine.
pagekit-first-page-decision	Step 03. Decide which page comes first — without defaulting to homepage.
pagekit-page-argument-shape	Step 04. Shape the argument for the chosen page.
pagekit-proof-map	Step 05. Map proof to sections. Name what cannot be said.
pagekit-first-page-draft	Step 06. Draft the page inside the argument shape and proof map.
pagekit-claim-check	Step 07. Adversarial audit at light / normal / hard severity.
pagekit-new-run	Scaffold a fully-logged run folder.
pagekit-run-check	Validate a run against the logging tiers.
pagekit-slop-check	Heuristic anti-slop regression against drafts.
pagekit-evaluator-pass	Invoke the evaluator-pass subagent on a completed run.

Two subagents (pagekit-claim-checker, pagekit-evaluator-pass) do the adversarial reads. The orchestrator dispatches to them automatically.

A worked example

examples/vegan-dog-food-verdel/ is the canonical run. Fully logged, non-homepage first page, claim-checked at severity: hard, evaluator pass included. Read first-page-draft.md against proof-map.md against claim-check.md to see how the chain constrains the draft.

examples/personal-crm-founders/ is a second worked example on a different object shape.

Without Claude Code

If you use ChatGPT, Perplexity, Grok, Gemini, or another chat window, open prompt-paste/ and paste the 7 prompts in order. Your output quality depends on how honest your source briefs are; that constraint does not change.

What makes this different

• Starts upstream, with signal, not with copy.
• Does not default to homepage. Step 03 forces you to name rejected candidates.
• Maps proof to sections. A section that cannot be proven is flagged or removed.
• Treats AI-slop patterns as hard no-go rules. pagekit-slop-check runs the regression; pagekit-claim-check catches the rest.
• Logs the whole run. The draft is not the deliverable; the logged run is.

Repo structure

• .claude-plugin/: plugin and marketplace manifests
• skills/: 12 SKILL.md files, each with a bundled references/ folder
• agents/: 2 subagents (claim-checker, evaluator-pass)
• examples/: worked runs (vegan-dog-food-verdel, personal-crm-founders)
• prompt-paste/: chat-app path with 7 prompts ready to paste
• runs/: your own runs land here (empty by default)
• scripts/: tooling (new-run, run-check, claim-check, slop-check, doctor)

License

MIT

Built by Hiten.