Skip to content

CLAUDE.md

Latest commit

 

History

History
166 lines (118 loc) · 7.85 KB

CLAUDE.md

File metadata and controls

166 lines (118 loc) · 7.85 KB

multai — Claude Code Instructions

Always adhere strictly to this file — it overrides all defaults.

0. Session Startup (Automatic)

At the very start of any new session, perform these steps automatically:

  1. Switch to Opus 4.6 (1M context) if not already selected.
  2. Read all project docs — this file and 100% of docs/.
  3. Compact the context — run /compact to free context for the task.
  4. Switch back to original model if it was changed in step 1.

Project Overview

1. Automated Enforcement

Seven layers enforce compliance:

  1. PostToolUse — Skill tracker — Records every Silver Bullet skill invocation
  2. PostToolUse — Stage enforcer — HARD STOP if quality gates incomplete before plan
  3. PostToolUse — Compliance status — Shows progress on every tool use
  4. PostToolUse — Completion audit — Blocks commit/push/deploy if required skills missing
  5. GSD workflow guard — Detects file edits made outside a /gsd:* command and warns
  6. GSD context monitor — Warns at ≤35% tokens remaining, escalates at ≤25%
  7. Redundant instructions + anti-rationalization — Workflow file + CLAUDE.md both enforce; explicit rules against skipping, combining, or implicitly covering steps

Trivial changes (typos, copy fixes, config tweaks): Automatically detected by hooks. Small edits (<300 chars) and non-logic files (.md, .txt, .css, .svg, etc.) skip enforcement per-edit. No action needed. Note: In CI/infrastructure changes (.yml, .yaml, .json, and .toml files), treat these as infrastructure code — they are NOT auto-exempted.

Subagent commits: Every git commit MUST use HEREDOC format and end with: Co-Authored-By: Claude Opus 4.6 noreply@anthropic.com

2. Active Workflow

The active workflow is loaded from docs/workflows/. Claude MUST read the active workflow file before starting any non-trivial task.

Active: docs/workflows/full-dev-cycle.md

Skill not found rule: If a skill listed in the workflow cannot be invoked, STOP and notify the user immediately. Do NOT silently skip.

2a. Browser Interaction

Always use mcp__Claude_in_Chrome__* tools for any browser/Chrome interaction. Never use mcp__computer-use__* tools for browser tasks — computer-use Chrome access is restricted and will be denied.

3. NON-NEGOTIABLE RULES

These rules apply to EVERY non-trivial change. There are NO exceptions.

You MUST NOT:

  • Skip a REQUIRED step because "it's simple enough"
  • Combine or implicitly cover steps ("I did code review while writing")
  • Claim a step is "not applicable" without explicit user approval
  • Proceed to the next phase before completing the current phase
  • Claim work is complete without running /gsd:verify-work

If you believe a step is genuinely not applicable, you MUST:

  1. State which step you want to skip
  2. State why
  3. Wait for explicit user approval before proceeding

"I already covered this" is NOT valid. Each Silver Bullet skill MUST be explicitly invoked via the Skill tool — implicit coverage does not count because the enforcement hooks track Skill tool invocations, not your judgment. GSD steps MUST be invoked as slash commands in the correct phase order.

Rules:

  • Do NOT stop until the final outcome is achieved
  • Always use /gsd:debug for ANY bug encountered during execution
  • Always use /forensics for root-cause investigation of completed sessions, abandoned sessions, or verification failures
  • CI must be green before deployment. Do NOT proceed with /deploy-checklist or /gsd:ship while CI is failing. Fix, re-push, re-check until green.
  • After every git push, ALWAYS wait for CI to complete and confirm green before declaring the task done. Poll with gh run view <run-id> until conclusion is success. If CI fails, fix the issue, re-push, and poll again. Never declare a push-dependent task complete without a confirmed green CI run.
  • README.md MUST be updated to reflect current version, features, and changes before release. /create-release will block if README is stale.
  • Pre-release quality gate is MANDATORY — see docs/pre-release-quality-gate.md. All four stages MUST pass (with /superpowers:verification-before-completion invoked per stage and markers recorded) before /create-release is invoked. No exceptions.
  • Always strictly adhere to this CLAUDE.md 100%

3a. Review Loop Enforcement

Every review loop (spec review, plan review, code review, verification) MUST iterate until the reviewer returns ✅ Approved TWICE IN A ROW. A single clean pass is not sufficient — the reviewer must find no issues on two consecutive passes. There are NO exceptions.

You MUST NOT:

  • Stop a review loop because "issues are minor"
  • Stop because "it's close enough"
  • Accept a partial fix and move on without re-dispatching
  • Count a round as approved unless the reviewer explicitly outputs ✅ Approved
  • Stop after a single clean pass — two consecutive clean passes are required

The loop is self-limiting: it ends when two consecutive clean passes are produced. If three or more consecutive iterations produce issues, surface to the user with the remaining issue list and wait for direction.

4. Session Mode

At the start of every session, before any work begins, ask:

Run this session interactively or autonomously?

  • Interactive (default) — I pause at decision points and phase gates
  • Autonomous — I drive start to finish and surface blockers at the end

Write the choice:

echo "interactive" > /tmp/.silver-bullet-mode
# or
echo "autonomous" > /tmp/.silver-bullet-mode

Fallback: if /tmp/.silver-bullet-mode is unreadable at any point, default to interactive and log "Mode fallback: defaulted to interactive" in the session log.

In autonomous mode:

  • Phase gates removed — proceed without approval pauses
  • Clarifying questions suppressed — make best-judgment calls, log each as "Autonomous decision"
  • Genuine blockers first (missing credentials, ambiguous destructive operations): these take precedence over all other rules — queue under "Needs human review", skip, surface in summary
  • Anti-stall (non-blocker stalls only): a stall = any of these three conditions:
    1. Same tool call with identical args producing the same result 2+ times consecutively
    2. 3+ tool calls in one step with no new state change (no file written, no decision, no new info)
    3. Per-step budget: >10 tool calls in one step AND no file written (Write/Edit resets counter) AND no autonomous decision logged since step began. Counter resets on Write/Edit, on any decision log event, and when a new /gsd: command or skill is invoked (new step boundary). On any stall: make best-judgment decision, move on, log under "Autonomous decisions".
  • All Agent Team dispatches use run_in_background: true
  • On completion: output structured summary (phases done, autonomous decisions, blockers queued, agents dispatched, commits made, virtual cost)

5. Model Routing

Default model: claude-sonnet-4-6 (latest Sonnet). No user friction for standard work.

Ask about Opus at two phase transitions only:

  1. Before Planning begins (before DISCUSS step):

    Entering Planning phase. Use Opus (claude-opus-4-6) for deeper reasoning, or stay on Sonnet?

  2. Before Design sub-steps apply (design-system / ux-copy / architecture / system-design):

    Entering Design phase. Use Opus, or stay on Sonnet?

If Opus permitted: switch to claude-opus-4-6 for that phase, return to Sonnet afterward.

Autonomous mode: stay Sonnet. Escalate silently to Opus only if a planning step produces measurably incomplete output: fewer than 5 lines, contains TBD/[TODO]/... placeholders, or a step expected to produce a file produces none. Log escalation as an autonomous decision.