Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/valternunez/refract/agents-mdgit clone --depth 1 https://github.com/valternunez/refractWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00988 | $0.00988 |
| Opus 5 | $0.00494 | $0.00494 |
| Sonnet 5 | $0.00198 | $0.00198 |
| Haiku 4.5 | $0.00099 | $0.00099 |
Grade A, and why
refract AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 66 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Refract — agent playbooks
Short recipes for when to reach for Refract and what to do with the output. Each is 5–15 lines: when to use it, the calls in order, what to look for.
render_responsiveis live and returns, in one response: structured findings (JSON keyed by preset), absolute saved paths, and a downscaled preview per viewport.
Playbook: verify a responsive bug fix
When: you changed CSS to fix a layout bug at a specific width.
render_responsive({ url: "http://localhost:3000/page", viewports: ["iphone-15", "ipad-mini", "macbook-air-13"] })- Look at the returned preview for the viewport you fixed.
- Confirm the
horizontal_overflow/element_clippedfinding for that viewport is gone from the response. If it's still present, the fix didn't take.
Playbook: visual-diff before merge
When: you want before/after proof a change didn't break the layout anywhere.
- On the base branch — MCP:
diff_responsive({ url, update: true }), or CLI:refract diff <url> --update --freeze— snapshots the baseline into./refract-baseline/{preset}.png. - Apply the change, then compare — MCP:
diff_responsive({ url })(returns a status per viewport + a diff image for each change), or CLI:refract diff <url> --freeze. unchangedeverywhere → safe to merge. Anychangedgives the % of pixels and a diff image/report.html(baseline │ current │ diff). The CLI exits non-zero on any change, so it drops straight into CI. Re-run withupdate/--updateto accept. Tip: passinjectCssto hide live clocks/ads so they don't read as false changes.
Playbook: screenshot a page behind a login
When: the page you need to verify is auth-gated (a dashboard, account settings).
- Get a Playwright storage-state JSON. If the project has e2e auth states (e.g.
e2e/.auth/*.json), use one of those. Otherwise a human runs once:npx playwright codegen --save-storage=auth.json <login-url>, logs in, closes it. render_responsive({ url: "https://app/dashboard", storageState: "./auth.json" })- The previews show the logged-in page (not the login wall); read findings as usual. If you still see the login screen, the state is stale — regenerate it.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 66 lines · 988 tokens per session scan A a1a7d7eb4805
refract AGENTS.md is an instructions file published in the GitHub repository valternunez/refract (0 stars, last pushed 2mo ago), licensed MIT. It adds 988 tokens to every session, about $0.0049 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
serve-sim AGENTS.md
Instructions for EvanBacon/serve-sim, covering native build notes, e2e testing with agent-browser and e2e testing via the serve-sim cli.
ironcurtain CLAUDE.md
Instructions for provos/ironcurtain, covering claude.md, general workflow, git & worktrees, git workflow and platform considerations.
ironcurtain security-boundary.instructions.md
Instructions for provos/ironcurtain, covering security boundary review rules, mandatory checks and structural invariant order.
ironcurtain sandbox.instructions.md
Instructions for provos/ironcurtain, covering sandbox review rules and mandatory checks.
ironcurtain pipeline.instructions.md
Instructions for provos/ironcurtain, covering pipeline review rules and mandatory checks.
ironcurtain types.instructions.md
Instructions for provos/ironcurtain, covering types review rules and mandatory checks.