Skip to main content

Team rules

Auto-generated. Do not edit by hand. Run pnpm capabilities or commit a rule change — scripts/build-capability-index.mjs regenerates this from .claude/rules/team/.

Team rules are conventions the whole team follows. A rule starts as a personal insight, is proposed via /share-rule (propose-team-rule), and once accepted applies in every session through @-imports in CLAUDE.md. Click a rule for its full text — the why, how to apply, and examples.

Accepted — in effect​

  • agent-drivable-cli — Every interactive CLI command must also have a non-interactive, machine-readable path (flags + --json) so an AI agent can drive it and read the result — not only a human at a TTY. (proposed by relbns, from sandbox)
  • artifact-language-english — User conversation may be Hebrew or English — match the user's language in chat. All written artifacts (docs, memory, GitHub, code comments, commit messages, Slack, ClickUp through tools) must be English only. (proposed by relbns, from bewith-docs)
  • confirm-before-acting — In any non-yolo flow, state a self-derived choice + rationale and wait for explicit approval before acting on it. (proposed by relbns, from bewith-docs)
  • explain-then-propose-fix — When the developer asks you to explain a problem, explain it then proactively propose a fix; on approval, present an implementation plan and get explicit approval before writing any code. (proposed by relbns, from bewith-docs)
  • inline-comments-critical-only — Keep inline code comments only when critical — non-obvious WHY, complex logic, or an approach that was already tried and failed. Move design rationale and history to the PR description or commit message body. (proposed by relbns, from leaders-app)
  • never-merge-without-approval — Claude must never merge a PR or branch silently. At the very least the developer must be AWARE of every merge Claude runs — surface which PR into which branch and get their go-ahead. A general "proceed"/"continue"/"ship it" about a task is not awareness of a specific merge. (proposed by relbns, from leaders-app)
  • never-offer-admin-merge — Claude must never proactively offer gh pr merge --admin (the branch-protection override). Report a blocked merge and leave the PR to a human; run --admin only if a developer explicitly instructs it. (proposed by relbns, from bewith-docs)
  • prefer-dedicated-tools-over-bash — Prefer dedicated tools over raw Bash — Read/Grep/Glob to read files, the sandbox infra-MCP query+write tools for DB/cache work — instead of cat/grep/find or raw mysql/psql/mongo/mongosh/redis-cli. (proposed by relbns, from bewith-docs)
  • session-completion-signal — When a session's planned work is done, tell the developer automatically (so they can close without losing context); proactively suggest a session handoff when context is long/stale or the next task is separable, and offer to draft a handoff prompt. (proposed by relbns, from bewith-docs)

Proposed — under discussion​

  • cache-key-by-viewer-permissions — A shared response cache on a permission-dependent endpoint must be keyed per viewer, or it serves one caller's data to another. (proposed by relbns, from backend-services)
  • concise-human-read-text — AI-authored text a human reads (PR bodies, commit bodies, ClickUp/Slack comments, review notes, reports) must be concise and lead with the point. (proposed by relbns, from backend-services)
  • dry-run-exercises-real-path — A dry-run must exercise or validate the same write / external-call path it claims to check, or explicitly log what it skips. (proposed by relbns, from migrations-tool)
  • integration-spec-ids-are-staging — Integration spec JSONs carry staging org/community/leader IDs; prod wiring must use the prod IDs, never the spec's. (proposed by relbns, from backend-services)
  • reusable-workflow-dynamic-matrix — Never drive a reusable-workflow (uses:) call from a matrix computed at runtime — it startup_fails the whole workflow; dispatch the target workflow instead. (proposed by relbns, from leaders-app)
  • test-at-the-layer-where-logic-runs — Logic that lives in the data layer (a MongoDB aggregation, a TypeORM repository / raw SQL query on Aurora-MySQL, or a DB transaction) must be covered by an integration test against the real engine — a DB-mocking unit test proves only the wiring, not the logic. (proposed by relbns, from backend-services)