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/shahriarshm/planchette/agents-mdgit clone --depth 1 https://github.com/shahriarshm/planchetteWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/shahriarshm/planchette/agents-md)<a href="https://agentmods.dev/instructions/shahriarshm/planchette/agents-md"><img src="https://agentmods.dev/badge/instructions/shahriarshm/planchette/agents-md.svg" alt="Measured on agentmods" height="20"></a>What 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 | $0.00647 | $0.00647 |
| Opus 5 | $0.00324 | $0.00324 |
| Sonnet 5 | $0.00129 | $0.00129 |
| Haiku 4.5 | $0.00065 | $0.00065 |
Grade A, and why
planchette 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 4d ago.
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 — 53 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Driving a window with the planchette CLI
You (the agent) are the control loop: capture the window, look at the screenshot, decide one action, run it, look again.
Setup
planchettemust be on PATH (uv tool install .from this repo).- macOS: the terminal app needs Screen Recording + Accessibility permission (System Settings → Privacy & Security), else captures are black and input does nothing.
- Windows: no special permission. Linux: X11 only (Wayland unsupported).
Selecting a window
- Interactive (macOS):
planchette pick— blocks until the user hovers and clicks a window under the overlay; allow a generous timeout (120s). - Headless / any OS:
planchette pick --listto see candidates, thenplanchette pick --name "Telegram"(case-insensitive substring of app name or title; frontmost match wins). planchette agentis the human-facing hotkey panel (it spawns its own agent session) — as an agent, never run it; usepick.
The loop
planchette capture→ printsPATH WIDTHxHEIGHT. View that PNG. All coordinates below are pixels in this image, origin top-left.- One action:
planchette click X Y(--doublefor double-click,--rightfor context menu)planchette drag X1 Y1 X2 Y2(select text, move sliders, drag & drop)planchette move X Yplanchette type "text"(types into the selected window — click the field first)planchette key "cmd+t"(cmd/ctrl/alt/shift + key;--repeat Nto press N times)planchette scroll X Y up|down [lines]
- Re-capture and look again. Repeat until done.
Text too small to read? planchette capture --crop X Y W H saves a native-res
zoom of that region (coords in last-capture pixels). It's a magnifier only —
keep computing click coordinates in the full capture's space.
Rules
- On macOS input goes to the selected window's process in the background — the user keeps their focus and cursor and can keep chatting while you drive. On Windows/Linux input follows real focus: warn the user not to touch anything.
- Never act blind: one action per capture, always look before the next.
- If actions have no visible effect in the next capture,
planchette raisethe window and retry once. error: no window selected/no capture yet→ run the missing precondition and retry.window is gone→ re-run pick.- If the window falls behind another,
planchette raise. - Stop and ask rather than guess or do anything destructive.
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.
- 4d ago First seen · 53 lines · 647 tokens per session scan A 0da589b8c912
planchette AGENTS.md is an instructions file published in the GitHub repository shahriarshm/planchette (0 stars, last pushed 2mo ago), licensed MIT. It adds 647 tokens to every session, about $0.0032 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
nanobrowser CLAUDE.md
Instructions for nanobrowser/nanobrowser, covering claude.md, project overview, development commands, workspace tips and architecture.
browser-control AGENTS.md
Instructions for keon/browser-control: Use this project as a tiny browser pipe, not an agent framework.
Upsonic CLAUDE.md
Claude Code instructions for Upsonic/Upsonic, covering claude.md, project overview, ai operational guides, default pre-work consultation and keep documents/ai/explanation/ in sync with code.
apple-browsers pixels.instructions.md
Instructions for duckduckgo/apple-browsers, a project described as: DuckDuckGo iOS & macOS browsers.
hunch-mcp AGENTS.md
AGENTS.md instructions for PrithviSeran/hunch-mcp, covering agents.md — working on hunch, setup & tests, architecture — one engine, three faces, adding or changing a tool and the focus-free contract (do not regress).
barebrowse CLAUDE.md
Claude Code instructions for hamr0/barebrowse, covering dev rules and project specifics.