Each agent is a markdown file under agents/ with its own system prompt, tool permissions, and model selection. Customize them. Fork them. Replace the ones you don't need.

The Workflow

           ┌─────────────┐
           │  Your Task  │
           └──────┬──────┘
                  │
         ┌────────▼────────┐
         │   📋 planner    │  ← Breaks into parallel subtasks
         │    (Tech Lead)   │    if touches 3+ files
         └────────┬────────┘
                  │
        ┌─────────┼─────────┐
        ▼         ▼         ▼
  ┌─────────┐ ┌─────────┐ ┌─────────┐
  │ fullstk │ │ fullstk │ │ fullstk │  ← Parallel execution
  │   × N   │ │   × N   │ │   × N   │    P7 methodology
  └────┬────┘ └────┬────┘ └────┬────┘
       └───────────┼───────────┘
                   ▼
           ┌──────────────┐
           │  🔍 critic   │  ← Mandatory pre-deploy review
           │   (reviewer)  │
           └───────┬──────┘
                   │
          ┌────────┴────────┐
          │                 │
          ▼                 ▼
   ┌──────────────┐  ┌─────────────┐
   │ 🐛 debugger  │  │   Deploy    │
   │ (if issues)  │  │             │
   └──────────────┘  └─────────────┘

Security-sensitive work takes a detour: critic flags → vuln-verifier writes PoC → fix or file PR.

The Methodology

Three Red Lines

Every agent enforces these. No exceptions. No "close enough".

• 🔒 Closure Discipline — Every task has a clear Definition of Done. "This is probably enough" is not an ending.
• 📎 Fact-Driven — Every judgment must cite actual code with paths and line numbers. "I guess" / "probably" / "should be" are violations.
• ✅ Exhaustiveness — Checklists cannot be skipped. Clean items must be explicitly marked "checked, no problems" — never silently ignored.

P7/P9/P10 Mode Switching

Not role-play. Operating modes that Claude switches between based on task scope:

| Scope | Mode | Behavior | |-------|------|----------| | Single feature | P7 (Senior Engineer) | Design → Impact analysis → Implement → Three-question self-review → [P7-COMPLETION] | | Multi-module, 3+ files | P9 (Tech Lead) | Decompose into Task Prompts with six elements. Coding is forbidden — your output is prompts, not code. | | Cross-team, 5+ sprints | P10 (CTO) | Output strategy docs. Goals, success metrics, risks, timeline, resource allocation. |

PUA Mode (High-Pressure Triggers)

The team shifts into exhaustive mode when:

• Same task failed 2+ times → write three new hypotheses, no retrying the old one
• About to say "I can't solve this" → forbidden, check docs and source
• Being passive, waiting for instructions → find the next step yourself
• User says "try harder" / "why did it fail again" → enter reflection mode
• User says "don't get slapped again" → cross-verify every assumption 3 different ways

We don't keep idle agents. No half-finished work. No excuses.

The Automation (Hooks)

Fifteen automation hooks wire up at pre-commit, post-tool-use, and stop events. They catch problems before they ship.

| Hook | Trigger | What it catches | |------|---------|-----------------| | 💰 cost-tracker.js | After every response | Token usage + estimated cost per model (Opus / Sonnet / Haiku). Running tally in ~/.claude/stats-cache.json | | ✋ commit-quality.js | Pre-commit | Blocks commits with debugger statements or hardcoded secrets in JS/TS/Python files | | 🔧 mcp-health.js | MCP tool failures | Detects MCP server outages and suggests restart paths | | 🛡 config-protection.js | Edit/Write to critical files | Guards important config files from accidental overwrites | | 🎨 design-quality.js | Frontend changes | Checks for AI-slop indicators in UI code | | 📝 check-console.js | Pre-commit | Flags stray console.log in production paths | | 📊 audit-log.js | All tool calls | Keeps an audit trail of significant tool operations | | 🎯 batch-format.js | Multi-file edits | Runs formatter on modified files in batch | | 💡 suggest-compact.js | Context pressure | Suggests /compact when context window fills up | | 📈 accumulator.js | Session tracking | Accumulates session metrics | | 🚨 log-error.sh | Any error | Unified error logging to ~/.claude/error-log.md | | 🧪 test-runner.js | After file edit | Finds sibling test file, runs vitest/jest, reports failures (non-blocking) | | 🔒 branch-protection.js | Pre-Bash | Hard-blocks force pushes and direct commits to main / master / production / release | | 📏 large-file-warner.js | Pre-Read | Warns at 500 KB, blocks at 2 MB to protect context window | | 📚 session-summary.js | Stop | Appends session summary to ~/.claude/sessions/ for later search |

Each hook is a self-contained script. Enable / disable / customize in settings.example.json.

From the Field

Real observations from daily use. Your mileage may vary.

critic is the MVP

On mid-sized modules (500–2000 lines), critic routinely finds 20–30 issues across all severity tiers. On large open-source codebases (tested against OpenClaw 352K⭐, Mermaid 87K⭐, Storybook 85K⭐, React Router 56K⭐), a single focused audit still surfaces 5–10 real bugs that had not been reported in the issue tracker.

Notable catches from real audits:

• A CWE-208 timing-safe comparison gap in a 352K-star repo that three prior security-hardening PRs had missed (diffs store !== vs safeEqualSecret)
• A non-atomic writeFileSync race in an auth-adjacent allowlist file that would corrupt state under concurrent access
• An Ollama reasoning-model heuristic regex (/r1/) that misidentified unrelated models as thinking models

The strictness ("assume everything is broken until proven otherwise") is what makes it work.

debugger saves you from face-plants

Twice during one bug-hunting session, the author was about to submit PRs based on seemingly clear reproductions. Both times debugger traced the behavior on HEAD and found the bugs had already been silently fixed in recent commits — the original reporters were on outdated versions. Submitting would have been embarrassing.

• Svelte #18083 — infinite-loop reconcile bug. Turned out to be a regression introduced in 5.43.8 that was fixed in 5.44.0+ by #17191 / #17240 / #1755