Claude Code Framework

How to Structure Claude Code
for GTM Operations

The folder architecture, governance layer, knowledge hierarchy, and execution system for running a full go-to-market stack with AI agents.

Governance
Paid Ads
Outbound
RevOps
Content
Analytics
Top-Level Architecture
The Workspace Root
Two core projects: one that governs, one that operates. Each is a separate Claude Code context with its own rules, skills, and agents.
gtm-workspace/
├── governance/                  ← The brain. Rules, audits, SOPs, blueprints
├── gtm-operations/              ← The muscle. Ads, outbound, RevOps, content, analytics
│
├── product/                     ← Optional: SaaS tools you're building
└── personal/                    ← Optional: private coaching, strategy, reflection
Every Project Has This Skeleton
Required Structure (per project)
Claude Code reads CLAUDE.md first, auto-loads rules by file path context, and activates skills when a task matches their description.
any-project/
├── CLAUDE.md                    ← Always loaded. Identity, critical rules, routing table
├── .claude/
│   ├── rules/                   ← Auto-loaded when working in matching file paths
│   │   └── *.md                    Guardrails, safety constraints, domain rules
│   ├── skills/                  ← On-demand. Triggered by task description matching
│   │   └── *.md                    Workflows, strategies, step-by-step procedures
│   └── agents/                  ← Subagents. Cheap models for research, review, QA
│       └── *.md                    Each agent = isolated context, single focused task
└── MEMORY.md                    ← Always loaded. Evolving state, quick references
1 · Governance — The Brain

Audit

Scan all projects,
flag deviations

→

Propose

File improvements
via /feedback

→

Apply

Human confirms,
changes go live

The governance principle: Working agents never modify skills, SOPs, or rules directly. They file proposals. Only governance reviews and applies changes after explicit human approval (CONFIRM). This prevents drift and keeps the system consistent.
governance/
Doesn't do client work. Governs how every project is structured, audits compliance, holds SOPs, and manages the self-improvement cycle.
governance/
├── CLAUDE.md                    ← Master rules. Analyze mode (default) vs Execute mode
├── .claude/
│   ├── rules/
│   │   └── governance-rules.md  ← Cross-project standards: naming, hierarchy, change control
│   ├── skills/
│   │   ├── audit.md             ← Scan all projects against blueprints, flag deviations
│   │   ├── session-end.md       ← Capture learnings before closing a session
│   │   └── review-proposals.md  ← Review improvement proposals filed by other projects
│   └── agents/
│       ├── researcher.md        ← Cheap model, returns concise summaries only
│       ├── code-reviewer.md     ← No parent context = completely unbiased review
│       └── qa-tester.md         ← Isolated test execution, keeps output clean
│
├── blueprints/                  ← One blueprint per project. Defines expected structure
│   ├── gtm-operations-blueprint.md
│   └── new-project-template.md
│
├── sops/                        ← Standard Operating Procedures (shared across all)
│   ├── skill-building-sop.md       How to build quality skills
│   ├── tool-building-sop.md        How to build composable tools
│   └── interaction-sop.md          How humans should use the system
│
├── registry/                    ← Living state that evolves over time
│   ├── learnings-log.md            What AI learned, tagged [PROPOSED] → [APPLIED]
│   ├── remaining-todos.md          Cross-project backlog tracker
│   └── proposals/                  Skill improvement proposals filed from working sessions
│
├── research/                    ← Decisions archive, audit history, research notes
└── checkpoints/                 ← Session state snapshots for context handoff
2 · GTM Operations — The Muscle

Knowledge

Strategy, methodology,
best practices per channel

→

Execution

API bots, scripts,
real operations

→

Context

Per-client data, ICPs,
history, deliverables

The 3-layer principle: Strategy never calls APIs. Bots never make strategic decisions. Client data stays isolated per client. Each layer has its own rules file that auto-loads based on which folder you're working in.
Paid Ads

LinkedIn, Meta,
Google Ads

Outbound

Cold email,
LinkedIn outreach

RevOps

CRM, pipeline,
lead routing

Content

Social, SEO,
thought leadership

Analytics

Attribution,
cross-channel

gtm-operations/
The root. Skills map 1:1 to GTM functions. Rules separate knowledge from execution from client data.
gtm-operations/
├── CLAUDE.md                    ← Three-layer routing: knowledge / execution / context
├── MASTER_SOP.md                ← Operating procedures specific to GTM work
│
├── .claude/
│   ├── rules/                   ← One rule file per architectural layer
│   │   ├── knowledge-rules.md      Strategy layer: methodology only, no live API calls
│   │   ├── execution-rules.md      Bot layer: API safety, credential handling, rate limits
│   │   └── clients-rules.md        Client layer: privacy, isolation, no cross-client data
│   │
│   ├── skills/                  ← One skill per GTM function / platform
│   │   ├── linkedin-ads.md         LinkedIn Ads strategy & campaign management
│   │   ├── meta-ads.md             Meta/Facebook Ads & audience targeting
│   │   ├── google-ads.md           Google Ads, keywords, search terms
│   │   ├── outbound.md             Cold email sequences & LinkedIn outreach
│   │   ├── revops.md               CRM ops, pipeline management, lead scoring
│   │   ├── content-social.md       Organic content, social, thought leadership
│   │   ├── client-audit.md         Full cross-channel account audit
│   │   └── report-builder.md       Automated report & doc generation
│   │
│   └── agents/
│       ├── researcher.md           Client research, competitor analysis, market data
│       └── code-reviewer.md        Review bot scripts for API safety & logic
│
│
├── ─── LAYER 1 · KNOWLEDGE ──────────────────────────────────────
│
├── knowledge/                   ← Strategy & methodology. Reads only, never calls APIs
│   ├── SKILL.md                    Master orchestrator — routes to the right subfolder
│   ├── CODE_OF_CONDUCT.md          Behavioral rules for all analytics, audits, reporting
│   │
│   ├── paid-acquisition/        ← Paid advertising strategy across platforms
│   │   ├── linkedin/
│   │   │   ├── SKILL.md
│   │   │   ├── knowledge-base/     ABM, audience sizing, bidding, formats, targeting
│   │   │   │   └── *.md            One file per topic (atomicity principle)
│   │   │   ├── methodology/        Your proprietary frameworks & approaches
│   │   │   └── templates/          Campaign plans, weekly logs, reporting structures
│   │   ├── meta/
│   │   │   ├── SKILL.md
│   │   │   └── knowledge-base/     Audience strategy, creatives, placements, scaling
│   │   ├── google/
│   │   │   ├── SKILL.md
│   │   │   └── knowledge-base/     Campaign types, keyword strategy, quality score
│   │   └── cross-platform/         Shared across all paid: personas, budgets, offers
│   │       ├── ad-personas.md
│   │       ├── budget-allocation.md
│   │       ├── channel-selection.md
│   │       ├── scaling-framework.md
│   │       └── measurement.md
│   │
│   ├── outbound/                ← Outbound prospecting strategy
│   │   ├── SKILL.md
│   │   ├── knowledge-base/
│   │   │   ├── cold-email-frameworks.md      Sequences, subject lines, CTAs
│   │   │   ├── linkedin-outreach.md         Connection requests, InMails, nurture
│   │   │   ├── icp-signal-mapping.md        Intent signals, trigger events, scoring
│   │   │   └── list-building.md             Enrichment, segmentation, hygiene
│   │   └── templates/              Sequence templates, cadence plans
│   │
│   ├── revops/                  ← Revenue operations strategy
│   │   ├── SKILL.md
│   │   └── knowledge-base/
│   │       ├── pipeline-stages.md           Stage definitions, exit criteria, SLAs
│   │       ├── lead-scoring.md              Scoring models, MQL/SQL criteria
│   │       ├── lead-routing.md              Assignment rules, round-robin, territory
│   │       ├── attribution-models.md        Multi-touch, first/last click, weighted
│   │       └── lifecycle-definitions.md     Lead → MQL → SQL → Opp → Customer
│   │
│   ├── content/                 ← Content & organic strategy
│   │   ├── SKILL.md
│   │   └── knowledge-base/
│   │       ├── content-pillars.md           Topics, formats, distribution channels
│   │       ├── social-playbook.md           LinkedIn organic, Twitter/X, engagement
│   │       ├── seo-strategy.md              Keywords, clusters, technical SEO
│   │       └── thought-leadership.md        POV development, authority building
│   │
│   └── analytics/               ← Cross-channel measurement & reporting
│       ├── SKILL.md
│       └── knowledge-base/
│           ├── kpi-definitions.md           North star, leading/lagging indicators
│           ├── reporting-cadence.md         Weekly, monthly, quarterly rhythms
│           ├── dashboard-specs.md           What to track, where, data sources
│           └── benchmarks.md                Industry/channel benchmarks by stage
│
│
├── ─── LAYER 2 · EXECUTION ──────────────────────────────────────
│
├── execution/                   ← API bots & scripts. Real operations happen here
│   │
│   ├── linkedin-bot/            ← LinkedIn Marketing API
│   │   ├── SKILL.md
│   │   ├── API_REFERENCE.md        Endpoints, auth, rate limits, gotchas
│   │   └── *.py                    Campaign CRUD, analytics, creatives, attribution
│   │
│   ├── meta-bot/                ← Meta Marketing API
│   │   ├── SKILL.md
│   │   ├── API_REFERENCE.md
│   │   └── *.py                    Audiences, campaigns, ad creation, placements
│   │
│   ├── google-ads-bot/          ← Google Ads API
│   │   ├── SKILL.md
│   │   └── *.py                    Campaigns, keywords, search terms, bidding
│   │
│   ├── email-sequencer-bot/     ← Outbound email tooling (Instantly, Smartlead, etc.)
│   │   ├── SKILL.md
│   │   ├── API_REFERENCE.md
│   │   └── *.py                    Sequence management, warmup, deliverability
│   │
│   ├── linkedin-outreach-bot/   ← LinkedIn outreach automation
│   │   ├── SKILL.md
│   │   └── *.py                    Connection requests, follow-ups, messaging
│   │
│   ├── enrichment-bot/          ← Data enrichment (Clay, Apollo, Clearbit, etc.)
│   │   ├── SKILL.md
│   │   └── *.py                    List building, contact enrichment, signals
│   │
│   ├── crm-bot/                 ← CRM integration (HubSpot, Salesforce, Attio, etc.)
│   │   ├── SKILL.md
│   │   ├── API_REFERENCE.md
│   │   └── *.py                    Contact sync, pipeline updates, deal tracking
│   │
│   ├── reporting-bot/           ← Automated report & doc generation
│   │   ├── SKILL.md
│   │   └── *.py                    Google Docs, Sheets, PDF generation, formatting
│   │
│   ├── tag-manager-bot/         ← Tracking & tag management (GTM, etc.)
│   │   ├── SKILL.md
│   │   └── *.py                    Tag creation, triggers, conversion tracking
│   │
│   └── analytics-bot/           ← Cross-platform unified analytics
│       ├── SKILL.md
│       └── *.py                    Pull data across all platforms, normalize, compare
│
│
├── ─── LAYER 3 · CLIENT CONTEXT ──────────────────────────────────
│
└── clients/                     ← Per-client data. Completely isolated from each other
    ├── client-a/
    │   ├── KNOWLEDGE_BASE.md       ICP, positioning, pain points, account IDs
    │   ├── LINKS.md                External links (Drive, dashboards, portals, Slack)
    │   ├── changelog.md            Running log of changes, decisions, actions
    │   ├── docs/                   Reports, proposals, deliverables, SOPs
    │   ├── notes/                  Call notes, meeting agendas, open items
    │   ├── audits/                 Account audit snapshots & findings
    │   └── scripts/                Client-specific automation scripts
    │
    ├── client-b/                   Same structure, same isolation
    └── client-c/
3 · Global Config — Safety & Shared Skills
~/.claude/
System-level settings that apply across every project. Hooks enforce safety before any tool runs. Global skills are available from anywhere.
~/.claude/
├── settings.json                ← Hooks, status line, permissions
│     hooks:
│       ├── PreToolUse         → Block destructive commands (rm -rf, force push, DROP)
│       ├── PreCompact          → Warn to checkpoint before auto-compaction
│       └── SessionStart        → Remind master rules after compaction
│
├── skills/                      ← Global skills (available from any project)
│   ├── checkpoint/
│   │   └── SKILL.md                Save session state before switching contexts
│   └── feedback/
│       └── SKILL.md                File improvement proposals from any project
│
└── projects/                    ← Per-project persistent memory (auto-managed)
    └── */memory/MEMORY.md          Evolving state, quick refs, survives across sessions
How It All Connects
Knowledge Hierarchy — What Gets Loaded When
Not everything loads at once. The system is tiered — critical context is always present, everything else loads only when needed.
Tier 1  │ CLAUDE.md                Always loaded. Identity + critical rules + routing.
Tier 2  │ .claude/rules/           Auto-loaded when you touch matching file paths.
Tier 3  │ .claude/skills/          On-demand. Activated when task matches skill description.
Tier 4  │ knowledge-base/ + SOPs   Referenced by skills. Deep methodology, how-to guides.
Tier 5  │ MEMORY.md                Always loaded. Quick lookups, evolving state.
Self-Improvement Feedback Loop
The system learns from every working session. Corrections and preferences get captured as structured proposals, reviewed weekly, and applied across the ecosystem.
Working session  ──→  /feedback  ──→  governance/registry/proposals/
                                              │
                                      /review-proposals  ←── weekly
                                              │
                                        Human approves  (CONFIRM)
                                              │
                                      Governance applies changes across all projects
                                              │
                                      registry/learnings-log.md  ←── captured
Change Control — What Can Be Modified and How
Not all files are equal. Constitutional files require explicit human approval. Working knowledge is freely mutable by the AI during sessions.
Constitutional   │ CLAUDE.md, SOPs, governance rules    ← Requires explicit CONFIRM
Structural       │ Skills, rules, agent definitions      ← Via /feedback → review cycle
Working          │ MEMORY.md, client data, learnings     ← Freely mutable during sessions