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/nyblnet/bento/agents-mdgit clone --depth 1 https://github.com/nyblnet/bentoWhat 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 | $0.01566 | $0.01566 |
| Opus 5 | $0.00783 | $0.00783 |
| Sonnet 5 | $0.00313 | $0.00313 |
| Haiku 4.5 | $0.00157 | $0.00157 |
Grade A, and why
bento 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 today.
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 — 111 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Working in this repo — agents & tools
Guidance for AI coding agents (Claude Code, Codex, Antigravity, …) and human
contributors. This file is the tool-agnostic contract; Claude Code also
reads CLAUDE.md (the deep architecture guide for slides/). If your tool
only reads one file, read this one, then follow the pointers.
What this project is
bento — office documents as single self-contained HTML files. One file = the
document + viewer + editor; it saves itself, updates itself over a signed
channel, and optionally syncs E2EE through a blind relay. slides/ is the
shipped app. Starting now: bento/spaces (Notion/notes-like),
bento/dash (spreadsheet + tables), bento/vault (document library).
Naming and casing — lowercase everywhere. The platform is bento, the
wordmark is bento/., and apps are bento/slides, bento/spaces,
bento/dash, bento/vault. This applies to UI strings and prose as well as
format constants — do not write "Bento Slides" in new copy. The / in the
wordmark is decorative: anywhere a name is stored or typed (filenames, URLs,
package names) it is plain bento. Full reasoning and the rejected candidates
are in docs/DECISIONS.md — don't reopen them.
Read before writing code
docs/PLATFORM.md— invariants every Bento app must honor. Breaking these bricks files already shipped to users.docs/PARALLEL-WORK.md— branch/merge discipline when many agents work at once (you are probably one of them).docs/DECISIONS.md— settled decisions. Don't relitigate them in code; append new ones.CLAUDE.md— deep architecture + hard-won gotchas, authoritative forslides/internals.docs/collab-design.md— the sync/collab spec + threat model.
Hard rules (each one has broken something before)
- Never let a literal
</script>into a bundle or document block. JSON in the doc block escapes<as<; builders concatenate around it. - The
#bento-docblock stays plaintext, same id, regex-extractable. That's the splice contract (docs/PLATFORM.md) — updaters already shipped in old files are frozen code that depends on it. - Never regenerate a document's
docId. It's the document's identity for recovery, sync, and future merge. - After any change to
slides/src/sync/crdt.ts, runnode scripts/test-sync.ts. The convergence rig has caught 15+ ordering bugs; a green typecheck means nothing for CRDT correctness. - A password-protected deck never carries a plaintext preview of page one.
Saves write a static first-page render into the shell for file-manager
thumbnails (
kernel/src/save.ts,slides/src/preview.ts);bento/encdecks are vetoed and any existing preview is stripped. Runnode scripts/test-preview.tsafter touching that path. - New UI strings go into ALL i18n catalogs. English-string-as-key; never
call
t()in module-level consts. Never take the locale list from prose — including this file. It isLOCALESin that app'sscripts/build-*-i18n.mjs, andls <app>/src/i18n/is the check. This rule said seven for a month after Portuguese made it eight, and #394 shipped an app one language short because it believed the sentence you are reading.- Adding a locale touches three places, not one: the catalogue file, the
app's
LOCALES, and any rig with its own copy of the list. Onlybuild-i18n.mjs(slides) errors on a catalogue missing fromLOCALES; the others exit 0 and silently never pack it. - A rig must derive the list, never restate it.
scripts/test-i18n-coverage.mjsreadsPACKED_LOCALESout of the generatedpacked.ts— copy that, and a fourth copy can never go stale. It covers slides ONLY (coreDiris hardcoded toslides/src/i18n), so spaces and type have no coverage rig at all. Do not assume your app's catalogues are checked by anything. - A locale code is not always a language.
ptis Brazilian inslides/and drifted European inspaces/; no check can see this, because both files are correctly namedpt.ts. See docs/DECISIONS.md, 2026-08-29, before writing or reusing a catalogue for a language with regional variants.
- Adding a locale touches three places, not one: the catalogue file, the
app's
- Never edit
site/— it's generated. Sources aresite-src/and thescripts/build-*.mjstooling. Same fordist-single/. - No AI co-author trailers on commits (no
Co-Authored-By: Claudeor similar), and no bot identities in git history. - Releases are cut locally by the maintainer only. Never touch signing
keys (
~/.bento/release-key.json), never attempt to release, publish, or deploy from an agent session unless the maintainer explicitly asks. - External PRs get provenance checks before merge (
gh api users/<login>) — AI-agent/bot contributions are not merged. - Verify before claiming done: typecheck, build, and exercise the change in a browser when it's user-visible. Report failures honestly.
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.
- today Changed · +19 lines · +340 tokens per session 483e9ab9e85c
- 3d ago First seen · 92 lines · 1,226 tokens per session scan A 59776bae22e8
bento AGENTS.md is an instructions file published in the GitHub repository nyblnet/bento (4,651 stars, last pushed yesterday), licensed MIT. It adds 1,566 tokens to every session, about $0.0078 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
ppt-master AGENTS.md
AGENTS.md instructions for hugohe3/ppt-master, covering agents.md, project overview, execution requirements, required conventions and compatibility boundary.
ppt-master CLAUDE.md
Claude Code instructions for hugohe3/ppt-master: Claude Code entry point. To avoid maintaining two copies, all project instructions live in a single source of truth — AGENTS.md — imported below (Claude Code inlines @-referenced files into context). Always edit AGENTS.md; never duplicate its content here.
genoffice CLAUDE.md
Claude Code instructions for genspark-ai/genoffice, covering claude.md, theming rules (mandatory) and build gotchas.
OpenAI4S CLAUDE.md
Instructions for PKU-YuanGroup/OpenAI4S, covering claude.md / agents.md, what this is, commands, the dual loop (the central architecture) and where things live.
macro CLAUDE.md
Claude Code instructions for macro-inc/macro, covering claude.md, architecture overview, key services, data storage and macrodb schema changes.
drizzle-cube CLAUDE.md
Instructions for cliftonc/drizzle-cube, covering claude.md, project structure, agent documentation context, database support and essential commands.