prose
AGENTS.md8 starsChanged 5 months agoMITView on GitHub
Shared instructions for OpenAI Codex and other non-Claude agents working in this repository. Claude-specific automation, slash commands, and skills live in CLAUDE.md and .claude/. When those files contain project architecture or workflow details that are not repeated here, treat them as additional repository guidance unless they conflict with higher-priority instructions from the user, system, or developer. Use the Trusted Coding profile (hz8WBqkgRrbiraXS9b32Sx) when working in this repo. Use the Review / Untrusted profile (2GV3WctmZUInztOStkCUMe) for read-only review passes. CLI shortcuts…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,700 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository31 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- accelerateSkill
- implement-issueSkill
- pipeline-engSkill
- pipeline-orchestrateSkill
- pipeline-peSkill
- pipeline-scorerSkill
- qa-acceleratedSkill
- qa-prSkill
- qa-webSkill
- releaseSkill
- review-feedbackSkill
- roadmap-prioritizationSkill
- roadmap-refinementSkill
- security-auditSkill
- accelerateSkill
- implement-issueSkill
- pipeline-engSkill
- pipeline-orchestrateSkill
- pipeline-peSkill
- pipeline-scorerSkill
- qa-acceleratedSkill
- qa-hitlSkill
- qa-prSkill
- qa-webSkill
- releaseSkill
- review-feedbackSkill
- roadmap-prioritizationSkill
- roadmap-refinementSkill
- screenshotsSkill
- security-auditSkill
The file
Show its rules and text
We split this AGENTS.md into the rules it gives an agent
27 rules. About 0 are common, 17 are distinctive, and 10 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈17
Distinctive · estimate
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 9
Use the Trusted Coding profile (hz8WBqkgRrbiraXS9b32Sx) when working in this repo.
- Very specific: line 10
Use the Review / Untrusted profile (2GV3WctmZUInztOStkCUMe) for read-only review passes.
- Distinctive: line 20
Follow system, developer, and direct user instructions first.
- Distinctive: line 21
Follow this file for cross-agent repository behavior.
- Distinctive: line 22
Follow CLAUDE.md for project architecture, Claude-specific workflows, and established local conventions.
- Very specific: line 23
Follow more specific docs under docs/ when working in a feature area they cover.
- Distinctive: line 25
If instructions conflict, choose the higher-priority source and mention the conflict in your handoff or PR.
- Distinctive: line 30
Before editing, check git status --short --branch and avoid overwriting unrelated changes.
- Distinctive: line 31
Do not revert user or other-agent changes unless explicitly asked.
- Distinctive: line 32
Keep edits scoped to the task.
- Distinctive: line 33
Prefer separate branches or git worktrees for independent tasks.
- Distinctive: line 34
Before comparing branches or opening PRs, run git fetch origin because another agent may have pushed recently.
- Very specific: line 35
Use gh CLI for GitHub operations.
- Distinctive: line 36
If you start a long-running process, make it clear how it is managed and do not leave required verification processes running …
- Distinctive: line 58
Use the repo's npm scripts instead of ad hoc command lines:
- Very specific: line 78
Cloud sandbox note: If npm run build fails at the build:skill step (requires zip), fall back to npx electron-vite build — it validates …
- Very specific: line 93
Avoid broad process kills such as pkill -f node or pkill -f Electron.
- Very specific: line 97
Keep platform-specific behavior behind getApi() from src/renderer/lib/browserApi.ts.
- Very specific: line 102
Do not use innerHTML with dynamic or LLM-provided content.
- Very specific: line 104
When changing IndexedDB stores in src/renderer/lib/persistence.ts, bump DB_VERSION and consider fresh-install and upgrade paths.
- Distinctive: line 105
Use shadcn/ui, Tailwind, and existing component patterns.
- Distinctive: line 106
Keep animations minimal and consistent with the existing app.
- Very specific: line 111
When adding architecture or workflow knowledge useful to future agents, update docs/ or this file instead of burying it in a PR comment.
- Distinctive: line 113
If you could not run a relevant check, state why.
- Distinctive: line 118
Keep one logical task per PR.
- Distinctive: line 119
Commit messages should reference the issue number when one exists.
- Distinctive: line 121
Before merging or declaring a PR ready, check unresolved review comments when applicable.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AGENTS.md
Shared instructions for OpenAI Codex and other non-Claude agents working in this repository.
Claude-specific automation, slash commands, and skills live in CLAUDE.md and .claude/. When those files contain project architecture or workflow details that are not repeated here, treat them as additional repository guidance unless they conflict with higher-priority instructions from the user, system, or developer.
Agent Profile
Use the Trusted Coding profile (hz8WBqkgRrbiraXS9b32Sx) when working in this repo.
Use the Review / Untrusted profile (2GV3WctmZUInztOStkCUMe) for read-only review passes.
CLI shortcuts (defined in ~/.zshrc):
oz-trusted --prompt "..."
oz-review --prompt "..."Instruction Precedence
- Follow system, developer, and direct user instructions first.
- Follow this file for cross-agent repository behavior.
- Follow
CLAUDE.mdfor project architecture, Claude-specific workflows, and established local conventions. - Follow more specific docs under
docs/when working in a feature area they cover.
If instructions conflict, choose the higher-priority source and mention the conflict in your handoff or PR.
Multi-Agent Coordination
- Assume Claude and OpenAI agents may work in this repo at the same time.
- Before editing, check
git status --short --branchand avoid overwriting unrelated changes. - Do not revert user or other-agent changes unless explicitly asked.
- Keep edits scoped to the task. Avoid opportunistic refactors.
- Prefer separate branches or git worktrees for independent tasks.
- Before comparing branches or opening PRs, run
git fetch originbecause another agent may have pushed recently. - Use
ghCLI for GitHub operations. Avoid interactive commands; pass flags explicitly. - If you start a long-running process, make it clear how it is managed and do not leave required verification processes running unintentionally.
Project Snapshot
Prose is an Electron desktop app with a React/TypeScript renderer and a shared code layer for main, preload, renderer, and MCP behavior.
Key areas:
src/main/- Electron main process, IPC, settings, integrations, updater, Sentry, MCP bridge.src/preload/- Context bridge for renderer-safe APIs.src/renderer/- React app, TipTap editor, Zustand stores, UI components. Chat tool-result renderers live insrc/renderer/components/chat/toolResultRenderers/(e.g.ListFilesResult.tsx).src/shared/- Shared schemas, tools, LLM utilities, and cross-process types.src/mcp-stdio/- Stdio MCP server used by Claude Desktop.docs/- Architecture notes, patterns, issue plans, troubleshooting, release docs, anddocs/roadmap.md(the canonical roadmap + operating model)..claude/- Claude Code settings, hooks, and skills. Do not treat these as Codex configuration.
Roadmap & Planning
Before any planning, prioritization, sequencing, or parallelization work — including "what to work on next," scoping a wave or epic, or splitting work across agents — read docs/roadmap.md first. It is the canonical, resumable source for what is being built, in what order (the wave model), and how work is parallelized (three concurrent Wave 1 tracks on an A→B→C merge ladder). The strategy already exists there; build on it rather than reinventing it. Project board #5 is the live queue; the roadmap is the narrative and operating model. It changes frequently, so read the current version — if your branch may be behind main, run git fetch origin and read git show origin/main:docs/roadmap.md rather than a possibly-stale worktree copy. Branch new worktrees off freshly-fetched origin/main.
Commands
Use the repo's npm scripts instead of ad hoc command lines:
npm run dev # Electron + Vite development server
npm run build # Production build to out/
npm run build:web # Web build
npm run test:unit # vitest unit tests (src/**/*.test.ts, jsdom)
npm run test:e2e # Build and run Electron Playwright tests
npm run test:web # Build and run web Playwright tests
npm run typecheck:e2e # Type-check E2E testsChoose the smallest verification that covers the change:
- Docs-only changes: no build required unless touching generated docs or examples that must compile.
- Renderer/main/preload/shared TypeScript changes: run
npm run build. - Web-specific changes: run
npm run build:webornpm run test:webwhen behavior changed. - Electron user flows: run
npm run test:e2ewhen practical, otherwise build and document manual verification. - E2E test changes: run
npm run typecheck:e2eand the relevant Playwright command.
Cloud sandbox note: If npm run build fails at the build:skill step (requires zip), fall back to
npx electron-vite build — it validates renderer/main/preload TypeScript. Do not use
npx tsc -p tsconfig.json --noEmit; it produces spurious project-reference errors (TS6305/TS6306) that
are not real type errors in this multi-tsconfig project.
Dev Server Safety
The Electron dev server writes .dev.pid. Prefer the PID file for this repo's process cleanup.
cat .dev.pid 2>/dev/null
ps -p <PID> -o pid=
kill <PID>Avoid broad process kills such as pkill -f node or pkill -f Electron; they can terminate another agent's session or MCP tools. If the PID file is missing, inspect processes carefully before killing anything.
Coding Conventions
- Keep platform-specific behavior behind
getApi()fromsrc/renderer/lib/browserApi.ts; do not accesswindow.apidirectly from renderer code. - Store settings through the existing settings paths and types. Do not introduce new
homedir()+.prosepaths for app data. - Preserve Electron sandbox settings:
contextIsolation: trueandnodeIntegration: false. - Classify every filesystem IPC path by source (comment
// Path source: A|B|C, seesrc/main/paths.ts): user-chosen →normalizePath()(a normalizer, not containment); content- or tool-derived →resolveWithin()with a containment root; typed names →sanitizeEntryName()in the renderer. - Store credentials through the OS-backed credential store; never commit plaintext secrets.
- Do not use
innerHTMLwith dynamic or LLM-provided content. - Gate unfinished features with
src/renderer/lib/featureFlags.ts; do not delete dormant feature code just to hide it. - When changing IndexedDB stores in
src/renderer/lib/persistence.ts, bumpDB_VERSIONand consider fresh-install and upgrade paths. - Use shadcn/ui, Tailwind, and existing component patterns. Do not add a new UI library without a strong reason.
- Keep animations minimal and consistent with the existing app.
Documentation And Handoffs
- For complex issue work, create or update
docs/issues/<number>/plan.md. - When adding architecture or workflow knowledge useful to future agents, update
docs/or this file instead of burying it in a PR comment. - In final handoffs, include what changed, what was verified, and any residual risk.
- If you could not run a relevant check, state why.
PR Expectations
- Work on a branch, not
main. - Keep one logical task per PR.
- Commit messages should reference the issue number when one exists.
- Include concise PR bodies with verification performed.
- Before merging or declaring a PR ready, check unresolved review comments when applicable.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

