Transform Claude into a proactive partner with the Proactive Brigade Kit for Claude Code (v2.5).

Proactive Brigade Kit for Claude Code (v2.5)

From reactive ассистент to proactive junior partner. Claude сам собирает бригаду экспертов, проверяет факты через веб и браузер, верифицирует смысл через профильного SME, и только потом показывает результат. Ты остаёшься CEO, не оператором.

What's new in v2.5

v1 → v2.5 — breaking redesign based on 10-agent meta-consilium.

Feature	v1	v2.5
Roles	9 permanent + dynamic	10 core + 57 on-demand (Domain Detection)
Trigger	User invokes `/master-agent`	Auto-trigger на medium+ через `disable-model-invocation: false`
Parallel spawn	Sequential опрос	Asymmetric context parallel spawn (anti-groupthink)
Fact verification	Text review	Fact-Checker agent с WebSearch (orthogonal)
Visual verification	❌	Visual QA через Claude Preview / Chrome MCP / Computer Use
Semantic review	❌	Universal dual-round debate (любой домен — медицина, юр, финансы, etc)
Smart routing	❌	Per-role model routing (haiku/sonnet/opus) + auto-escalate
Private data protection	❌	Hard-filter + sanitizer перед spawn для `*_private.md`
Error paths	❌	Spawn fail fallback, recursion cap, TTL вердикта, hotfix bypass, interrupt
Arithmetic check	❌	Self-check 3× rule
Rollback	❌	Git snapshot + tag-based rollback

См. CHANGELOG.md для полного списка breaking changes.

Demo — реальный консилиум

You: "Посчитай ROI для промислової СЕС 500кВт под tariff 2026"

Claude: Вижу medium задачу (числа + регуляторика + потенциально публичный контент).
        Предлагаю консилиум (15 сек):
          ⚡ PV Engineer → sonnet — tech specs
          ⚡ Finance Engineer → opus — CAPEX/OPEX benchmarks + methodology
          ⚡ Regulatory Expert → opus — tariff model 2026
          ⚡ Fact-Checker → opus (WebSearch) — 5 key numbers
          ⚡ Critic → opus — adversarial assumptions
        Fresh Data: tariffs/rates/CAPEX 2026
        Запускаю? [y/n/custom]

You: y

[5 agents параллельно, ~2 min]

Claude: CONSILIUM VERDICT:
        🎯 РЕШЕНИЕ: Base-case payback 7.3 года, range 5-11 years
        🔴 3 BLOCKERS:
          B1: «зелений тариф» для 500kW NOT AVAILABLE (Fact-Checker caught
              via WebSearch — outdated assumption from training)
          B2: War-risk insurance — critical CAPEX/OPEX missing (Critic flag)
          B3: Curtailment 5-15% mandatory (Critic flag)
        🎯 CONFIDENCE: 52% (will increase после B1-B3 решения)
        📎 Logged to decisions/2026-04-19_ses-500kw-roi/

Without v2.5: reactive Claude would give «4-5 years» confidently using outdated tariff assumption → fake numbers to client. With v2.5: caught outdated regulation via WebSearch, flagged war-risk insurance, arithmetic self-checked 3×.

Quick Start

1. Clone / download

git clone https://github.com/3952801-sudo/claude-code-master-agent.git
cd claude-code-master-agent

2. Copy skills to Claude Code config

# Windows (Git Bash) / Linux / macOS
cp -r skills/master-agent ~/.claude/skills/
cp -r skills/consilium ~/.claude/skills/
cp -r skills/role-generator ~/.claude/skills/

3. Copy templates and customize

cp CLAUDE.md.example ~/.claude/CLAUDE.md
cp triggers.md.example ~/.claude/triggers.md

Edit ~/.claude/CLAUDE.md and ~/.claude/triggers.md:

Replace placeholder company/project names
Add your industry-specific triggers (see bottom of triggers.md.example)
Customize your fleet of SMEs in ~/.claude/skills/role-generator/SKILL.md "Industry-Specific Experts" section

4. (Optional) Set up private memory protection

# Your sensitive project memory
mkdir -p ~/.claude/projects/your-project/memory
# Name sensitive files with _private suffix — hard-filtered by consilium
touch ~/.claude/projects/your-project/memory/project_nda_thing_private.md

5. Start using

Open Claude Code. For any medium+ task, Claude will automatically offer brigade consilium.

Kill-switches if needed:

/quick — next task trivial mode
/no-brigade — disable proactivity for session
/consilium [roles] — force consilium
/restore — re-enable proactivity

Architecture

Claude Code main session (orchestrator)
   │
   ├─ Trigger match → determine class (trivial/small/medium/large)
   │
   ├─ trivial/small → direct work + inline Critic
   │
   └─ medium/large → Proactive Offer → [y] spawn parallel Agent:
        │
        ├─ SME role A (asymmetric context: only tech specs)
        ├─ SME role B (asymmetric context: only regulatory docs)
        ├─ Copywriter (sees tech brief)
        ├─ Critic (sees ONLY outputs, NOT task brief — adversarial)
        ├─ Fact-Checker (WebSearch, verifies numbers)
        │                                      ↓
        │            [15-30 sec parallel]
        │                                      ↓
        └─ Synthesize 300-word verdict with max 3 blockers + confidence
           → Log to ~/.claude/decisions/<task>/
           → Continue work with вердикт
           → Post-work: Visual QA / Semantic Review / Final Auditor
           → Show to user

What's inside

skills/
├── master-agent/          Orchestrator skill — always use for medium+
│   ├── SKILL.md           Strong description, auto-trigger
│   └── references/        10 core roles, 110 brigade rules, 10 workflows
├── consilium/             Parallel spawn playbook — the core pattern
│   ├── SKILL.md           Asymmetric context, veto hierarchy, error paths,
│   │                      Visual Verify, Semantic Review Protocol,
│   │                      Smart Routing Matrix, Tool Gating Matrix
│   └── references/
│       └── cutover-smoke-test.md   dev→prod playbook
└── role-generator/        67 templates + meta-template for any SME
    └── SKILL.md           Cross-domain SME examples, customize per industry

CLAUDE.md.example          Global rules template
triggers.md.example        Regex matrix template (two-phase triggers)
CHANGELOG.md               v1 → v2.5 breaking changes
LICENSE                    MIT

Six Axes of Quality

Each axis has dedicated roles + its own orthogonal verifier:

Axis	Roles	Orthogonal Verifier
🔢 Technical numbers	Domain SME	Fact-Checker with WebSearch
⚖️ Legal risks	Legal Compliance	Critic (hard veto)
🎨 Design	UX + Visual + Accessibility + Creative Director	Visual QA (real screenshots + contrast)
✍️ Text content	Content Strategist + Copywriter + Editor + Domain SME	Readability + Critic + SME Semantic Review round 2
🔍 SEO	SEO Specialist + Technical SEO + Local SEO	SEO Audit via WebFetch
🧠 Psychology / CRO / Trust	Sales Psychologist + CRO + Trust Expert	UX Flow Tester via Chrome MCP

Key Patterns (SOTA multi-agent)

Asymmetric Context (anti-groupthink)

7 copies of same Claude with same prompt = 1 opinion × 7. Role-prompting alone is cosmetic. Solution: each role sees only its relevant slice of context, not the full brief. Critic specifically sees ONLY outputs of others (no task brief) to stay adversarial.

Orthogonal Verification (anti-hallucination cascade)

If PV Engineer hallucinates a number and Critic has the same context, Critic confirms the hallucination. Solution: Fact-Checker is a separate spawn with WebSearch + ONLY the claims to verify (no task context). Similarly, Visual QA sees only the rendered page (no source code).

Universal Semantic Review

Any text with domain topic gets dual-round debate:

Round 1: Domain SME writes tech brief → Copywriter drafts
Round 2: Same SME reads draft, marks each claim ✓/⚠️/🔴
Round 3: Copywriter revises per SME notes
Exit: SME approves OR 2 iterations → escalate to owner

Smart Routing per role

haiku  — classifiers, triggers, simple fact-check (versions, dates)
sonnet — specialist work (Copywriter, Visual QA, UX, most SMEs)
opus   — adversarial (Critic, Auditor), judgment (Legal, Semantic Review),
         multi-step (Architect, Analyst)

Hard-filter for NDA data

Files with suffix *_private.md или marker DO NOT SHARE are physically excluded from spawn context. Before spawn, orchestrator greps and either excludes the file or sanitizes task descriptors (replaces {client} → ClientA, {location} → RegionX).

Known limitations

Not a replacement for real experts — for high-stakes medical/legal/financial decisions, always get real 2nd opinion. Semantic Review catches obvious errors but isn't a licensed professional.
Cost: 5-10 parallel opus spawns per medium task ≈ $0.30-1.00 per consilium. Use /no-brigade for quick prototyping. Daily budget alert recommended.
No live Chrome in every environment — Visual QA requires Claude in Chrome MCP or Claude Preview. Fallback via mcp__Claude_Preview__* for local dev servers.
Regulatory updates require WebSearch — Fact-Checker must have WebSearch tool access explicitly (see tool gating matrix in consilium SKILL.md).
Universe of domain SMEs is meta-templated, not exhaustive — for a new industry, extend role-generator/SKILL.md "Industry-Specific Experts" section.

Migration from v1

If you were using master-agent v1 (9 roles, no Critic, no consilium):

Backup your v1: mv ~/.claude/skills/master-agent ~/.claude/skills/master-agent.v1-backup
Install v2.5: see Quick Start above
Migrate your private memory: rename project_company.md → project_company_private.md, add sensitivity: PRIVATE frontmatter
Update CLAUDE.md: use CLAUDE.md.example as starting point (v1 didn't have global CLAUDE.md)

See CHANGELOG.md for full v1→v2.5 delta.

Contribute

PRs welcome, especially:

New industry SME templates (manufacturing, healthcare, construction, logistics, etc)
Language-specific triggers (currently most patterns support EN/RU/UA mixed)
Additional verification tool integrations
Example decisions/ verdicts from your production work

License

MIT — see LICENSE

Credits

Developed through 10-agent meta-consilium reviewing own architecture (April 2026)
SOTA patterns inspired by: CrewAI (asymmetric roles), AutoGen (GroupChat debate), Magentic-One (orchestrator ledger), LangGraph (state machine)
Extracted from real B2B EPC (Engineering-Procurement-Construction) workflow