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/mirror4869/material-workbench/agents-mdgit clone --depth 1 https://github.com/mirror4869/material-workbenchWhat 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 | $0.01140 | $0.01140 |
| Opus 5 | $0.00570 | $0.00570 |
| Sonnet 5 | $0.00228 | $0.00228 |
| Haiku 4.5 | $0.00114 | $0.00114 |
Grade A, and why
material-workbench 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 2d 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 — 87 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md
Project scope
Material Workbench is a Windows local-first media catalogue. Source code and static assets live in this repository; real media, user metadata, generated indexes, thumbnails, journals, logs, and local configuration do not.
Safety rules
- Never use a real media library in tests. Use temporary directories and synthetic fixtures.
- Never commit or directly edit
user_data.jsonor runtime journals. - Existing library files must not be silently moved, renamed, overwritten, or deleted.
- Desktop intake defaults to a no-overwrite library copy. Moving the original is an explicit warned choice; a copy does not silently relink an external project or make its Desktop source safe to delete.
- Destructive or transactional physical changes require preview, explicit confirmation, authoritative server-side revalidation, persistent recovery evidence, and rollback. First-ingest copy instead keeps the source, publishes atomically without overwrite, and removes only its own failed target.
- Destructive and Adobe write APIs accept opaque IDs, not browser-provided filesystem paths.
- Do not weaken reparse-point, hard-link, stale-plan, project-identity, or concurrent-write guards to make a test pass.
- Adobe integration is optional and must fail closed when the external bridge or same-project runtime guard is unavailable.
- Keep all text files UTF-8 and preserve stable asset ID algorithms.
- Core workflows must remain usable without AI. Agent output is advisory and cannot bypass user confirmation, server-side validation, or recovery gates.
- Browse and Organize are presentations of one library. New item actions and query concepts should share a command/query contract instead of creating a capability that exists in only one presentation.
- Premiere, Blender, DaVinci Resolve, CapCut/Jianying, and other creative applications are optional capability-declared connectors. Keep reveal and copy-path usable without a connector, and never simulate a host write with UI automation or expose an action whose runtime guard is unavailable.
- Keep
assets/js/workbench-query.jspure and explicitly driven by item, query, and context inputs. The browser uses catalogue → query → bootstrap → oracle → main order; Organize/Material is the only migrated query adapter. Keep Browse/Home on its legacy implementation until independent equivalence tests pass, and do not imply that both interfaces already share all semantics. /api/bootstrap/statusmust remain read-only: it may diagnose but must not scan, repair, install, or write.assets/js/workbench-bootstrap.jsmay synthesize an empty payload only for a source explicitly reported asmissing;invalidorunreadabledata blocks loading. Optional dependency failures must gate only the capabilities that require them, and every gated action must also fail closed inside its function body.- Keep
inspiration_cards_core.pyindependent of the HTTP server. Inspiration registry v1 reads are zero-write; an explicit first v2 write requires its dedicated verified backup and one atomic save with one overall revision increment. V2 cards allow 0 to 6 ordered references, at most one cover in the first position, server-derived labels for local IDs, and URLs that the server records without fetching. Migration must inspect every local media/document reference in both schema versions, and exact deduplication must inspect every local asset reference. - The visible inspiration editor supports 0 to 6 ordered references. Keep v1
reads and text-only edits zero-upgrade; send an explicit v2 request only
after a real add, remove, reorder, or cover change. Local media and documents
must come from the catalogue picker, never from a browser-provided ID or
path. If the backend does not explicitly advertise
multi_ref_v2, every rich-reference mutation and v2 save must fail closed while compatible v1 text editing remains available. - Ordinary Recycle Bin deletion must check every matching v1/v2 local card reference and confirm it against the current inspiration-registry revision. Keep that confirmation independent from Adobe dependency confirmation; a drift in either revision performs no file move or metadata write. After a confirmed deletion, preserve the card and stored reference and resolve the source as missing.
- Keep Daily Oracle local, optional, and non-authoritative. It may open only
after an explicit user click and must never be restored as an open dialog by
session recovery. The selected deck and normally browser-random salt belong
only in
localStorage; a fixed non-identifying fallback is allowed when random generation or storage is unavailable. None of these values may enteruser_data.json, an index, or a server API. A draw stays stable for the user's local calendar day. Present the bundled traditional and original decks as entertainment and self-reflection, never as medical, legal, financial, safety, or other professional advice. Traditional content may use public-domain Major Arcana symbolism, but must not copy a modern deck's art, text, brand, or interactive assets.
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.
- 2d ago First seen · 87 lines · 1,140 tokens per session scan A 627af2c07e78
material-workbench AGENTS.md is an instructions file published in the GitHub repository mirror4869/material-workbench (99 stars, last pushed 17d ago), licensed Apache-2.0. It adds 1,140 tokens to every session, about $0.0057 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 2026-08-30.
Other instructions, from other repositories
mosa AGENTS.md
AGENTS.md instructions for fengseekling-coder/mosa, covering mosa contributor instructions, repository boundary, local development, licensing and safety.
appwrite AGENTS.md
AGENTS.md instructions for appwrite/appwrite, covering appwrite, commands, stack, layout and libraries.
siyuan AGENTS.md
AGENTS.md instructions for siyuan-note/siyuan, covering agents.md, 1. non-negotiable constraints, do not hand-edit, verification and prohibited operations and 2. project-specific rules.
veritas-kanban AGENTS.md
Instructions for BradGroux/veritas-kanban, covering agents.md — canonical agent instructions for veritas kanban, runtime requirements, repository layout, essential commands and install.
ai AGENTS.md
AGENTS.md instructions for vercel/ai, covering agents.md, project overview, repository structure, key directories and core package dependencies.
stackql CLAUDE.md
Instructions for stackql/stackql, covering claude.md and claude code specifics.