Make the agent install prompt refusal-resistant: verify-first, steps inlined

[ejntaylor]

Summary

Iterates on the AI-agent install prompt after field-testing three distinct refusal modes:

1. Remote-doc delegation ("follow https://patchstack.com/install.txt") reads as remote script execution — agents refuse before fetching (fixed in Ship an agent-facing install guide inside the npm package #38).
2. Knowledge-cutoff hallucination — agents whose training predates May 2026 assert @patchstack/connect doesn't exist; the prompt now includes a read-only npm view @patchstack/connect check so the registry, not model memory, settles existence.
3. In-package delegation + persuasion detection — "install the package, then follow the instructions it ships" is structurally prompt-injection-shaped, and reassurance language ("published after your cutoff, don't conclude it doesn't exist") is itself flagged as a manipulation signal.

The README prompt is now five numbered, individually inspectable commands — no delegation, no trust arguments. The bundled AGENT-INSTALL.md / npx @patchstack/connect guide stays as reference material (framework-specific widget placement, re-run rules, troubleshooting) rather than the instruction channel the install depends on, and its knowledge-cutoff note is rephrased to be purely factual.

🤖 Generated with Claude Code

[coderbuds]

Clear, focused addition of a verify-first step to the install prompts.

🎯 Quality: 98% Elite · 📦 Size: Tiny

📈 This month: Your 217th PR — above team average · Averaging Excellent

_{See how your team is trending →}