Quote Card

A generated HTML plan is ~79KB of CSS, inline SVG, and JavaScript. The actual spec inside it — objective, steps, tests, acceptance criteria — is a few KB. So every time an AI agent re-opens that plan to implement, review, or fan it into a workflow, it burns ~20K tokens of presentation noise to recover what it needs.

The fix is a pattern worth stealing: embed a machine-readable digest.

In plan-agent's HTML plans, the first child of <body> is a <script type="text/markdown" id="plan-digest"> block. Three things make it work:

1. It's inert. Browsers don't render or execute an unknown script type, so the plan looks and behaves exactly the same — zero visual or runtime cost.

2. It's spec-only. The digest carries objective, context, files, steps (action/why/verify), tests, and criteria — and omits all mutable state. Status, checkboxes, and progress live in <head> meta tags and the live DOM, never in the digest. So flipping a status can't corrupt the spec.

3. It's cheaply extractable. One awk line pulls just the spec — downstream tools read ~79 lines instead of ~79KB.

The principle: progressive disclosure applied to a document. Ship the human render and the machine spec in the same file, and let each reader take only the layer it needs.

More Claude Code plugin patterns like this on my feed.

#ClaudeCode #AITools #DeveloperTools #Automation

Twitter/X

AI agents re-read your HTML artifacts constantly. Don't make them parse 79KB of CSS/JS for a few KB of spec.

plan-agent embeds a spec-only <script type="text/markdown"> digest as <body>'s first child — inert, state-free, one awk line to extract. #ClaudeCode

Bluesky

Nice pattern from agentics' plan-agent: its HTML plans embed a spec-only digest as <body>'s first child via <script type="text/markdown">.

The file is ~79KB of CSS/JS; the digest is the few-KB spec an agent needs — inert, state-free, one awk line to pull. Progressive disclosure for a doc.

Substack

A small idea with outsized payoff for agent workflows.

plan-agent's HTML plans are ~79KB of styling and script — but the spec an agent needs (objective, steps, tests, criteria) is a few KB. So each plan embeds a spec-only <script type="text/markdown"> digest as <body>'s first child: it never renders, holds no mutable state, and extracts with one awk line.

When an artifact has a human and a machine reader, give each its own layer in one file.

More Claude Code patterns in the newsletter.