write-for-agents

Content Readiness Kit

Your content has a second reader now. This is a kit for writing for it.

When someone decides whether to use what you build, they increasingly don't read your page first - they ask an AI, and act on the three-sentence summary it hands back. That AI reads you before any human does, compresses you, and forms an opinion. If your content is generic, it gets averaged into the category and you disappear. This kit is a short, practical discipline for making your content survive that reading - without losing the human reader.

Think of it as an ATS for everything you publish. An applicant tracking system reads your resume before a recruiter does and decides whether a human ever sees you. The same thing now happens to all your external content - docs, product pages, posts, launches: an AI reads it before your buyer does, compresses it, and decides whether they ever hear your name. With one decisive difference. You don't beat this reader by stuffing keywords. It's a reasoning model that checks your claims, so you win by being provably specific, not by gaming it. (Trying to game it backfires - if the agent can't verify a claim, it just drops it.)

It's prose and templates, not software. No install, no dependency. Read it in 15 minutes; apply it the same day.

---

What's in here

File	What it is
content-readiness-rubric.md	A scored 5-dimension rubric for whether a piece of content is agent-legible (provable claims, claim-evidence pairing, compression survival, honest wedge, machine-clean structure).
two-audiences-explainer.md	The one-page "why" - the two readers, compression, the truth layer, the honest wedge. Read this first.
how-to-apply.md	How to run the rubric in practice: where it fits, the compression test, integrating into review without a bottleneck.
compression-test-template.md	A drop-in worksheet: paste your content into an AI, see whether your differentiator survives compression.
worked-example.md	A generic before/after: the same announcement scored 4/10, then 10/10, with the three fixes shown.
examples/	The kit applied to real public articles (Cloudflare 10/10, Ethereum 7/10, Vercel 6/10) - field notes with compression tests and the lift for each.

What you can do with it

• Audit any external-facing content - a product page, a doc, a README, a blog post, a launch announcement, even your own LinkedIn - for whether an AI agent can extract a real, provable, differentiated point from it.
• Run the compression test to see, in two minutes, what an AI tells a buyer about you when it reads your page. (Spoiler: most pages average into the category, and it's invisible from the inside.)
• Add it to your review checklist so content gets an agent-legibility pass before it ships, the way code gets a test pass before it merges.

How to use it in your own workflow

Pick the tier that fits how you work:

1. Solo / ad-hoc. Before you publish anything external, run the compression test on it. If your differentiator doesn't survive, fix that one thing. That's the whole habit.
2. In review. Drop the 5-dimension rubric into your content-review or docs-PR checklist. A failing dimension blocks publish the way a failing test blocks merge.
3. At generation time. If you draft with an AI assistant, paste the five checks into your prompt so the first draft is already agent-legible. Highest leverage - you fix the cause, not the symptom.

It works the same whether you're in eng, product, docs, devrel, or marketing. If you write something an outsider will read to make a decision, it has a second reader, and this kit is for serving it.

Using Claude Code? The repo doubles as a plugin: run /plugin marketplace add kpachhai/write-for-agents, then /plugin install write-for-agents@write-for-agents. The content-audit skill runs the rubric scoring + compression test on any content you point it at, calibrated against the kit's worked examples.

Running a content pipeline? If your team has an editorial workflow - or an AI-assisted drafting process - this is the agent-readiness gate that sits alongside your editorial, plain-language, and compliance checks: one more pass asking "does our differentiator survive when an AI summarizes this?" It pairs with voice, it doesn't replace it. Keeping the author's voice serves the human reader; the truth layer serves the agent. You need both - that's the dual-channel discipline.

The 10-minute quickstart