SKILL.md is the install path

Your next buyer may never land on the marketing site. An agent hits the repo, reads the instruction file, tries the skill, and either installs you or routes around you. You will not see that miss in analytics.

Anuraag put it plainly: the agent does not read the landing page — it hits a wall, or it routes around you.

What changed

Repositories now ship instruction files for coding agents: CLAUDE.md, AGENTS.md, .cursor/rules, and SKILL.md / skill packs. That is not a docs nicety. It is the first screen the agent sees.

agentcheck on DEV (17 Oct): those files are now part of how the stack boots. If they are wrong, stale, or missing, the agent fails the five-minute test and moves on.

Waldek on DevAX: developers judge your technology by what their agent builds. In a follow-up, a skill alone scored poorly; after fixing two docs the agent already read, the same path jumped to 83/85 without the skill. The install surface was the docs the agent already touched — not the homepage.

Skill = install path, not a glossary

A useful skill does one job well: enough context for the agent to finish a real task with your product. It is closer to a working example than to a feature list.

What to ship so an agent can install you:

1. One instruction file that names the product — what it is for, what it must not do, where the canonical docs live.
2. A skill (or rules pack) that completes one workflow — install, auth, first successful call. Not ten.
3. Docs the agent already opens — markdown that answers the failure modes the skill hits. Fix those before you add another landing section.
4. A tested path — Josh Crites: llms.txt and markdown docs, an MCP server, skills/rules, tested examples. The checklist is boring on purpose.

If your GTM still treats the homepage as the top of funnel and the repo as "support," you are optimizing the page humans share in Slack while the agent never loads it.

What this is not

• Not another "what is a skill" explainer. GitHub and Claude Academy already own that SERP.
• Not a replacement for GEO. Citation in ChatGPT and install by Claude Code are different jobs (see Cited ≠ installed).
• Not a promise that shipping SKILL.mdalone moves pipeline. Waldek's numbers went up when the docs the agent already read got fixed.

How GTM Labs uses this

Inside a Growth Pilot, agent experience is a module: make the install path finish, measure whether an agent can complete the first task, then keep or kill on Friday. Same discipline as the rest of the pilot — one experiment card, a number, a re-run.

Related: AI Agents playbook · Go-to-market for AI agents · Cited ≠ installed