claude-code-for-agentic-gtm
This repository packages 12 Claude Code skills for B2B GTM teams that want to move from isolated AI prompts to a governed workflow. It turns market changes into sourced account research, commercial decisions, approved actions, CRM records, reply analysis, and reusable GTM learning.
Claude Code for Agentic GTM
This repository packages 12 Claude Code skills for B2B GTM teams that want to move from isolated AI prompts to a governed workflow. It turns market changes into sourced account research, commercial decisions, approved actions, CRM records, reply analysis, and reusable GTM learning.
What's Inside
| File | What it does |
|------|-------------|
| .claude/skills/agentic-gtm/SKILL.md | Orchestrates the complete 12-stage workflow and preserves run state. |
| .claude/skills/change-detector/SKILL.md | Converts supplied source material into dated candidate changes without inferring intent. |
| .claude/skills/evidence-gatherer/SKILL.md | Validates claims, records contradictions, and preserves source evidence. |
| .claude/skills/account-researcher/SKILL.md | Builds a bounded account brief from supplied research questions. |
| .claude/skills/commercial-relevance-assessor/SKILL.md | Scores fit, timing, evidence quality, and reachability using a fixed rubric. |
| .claude/skills/why-now-hypothesis-builder/SKILL.md | Builds a falsifiable commercial hypothesis without presenting it as fact. |
| .claude/skills/next-action-selector/SKILL.md | Selects a reversible next action using evidence, policy, and relationship context. |
| .claude/skills/evidence-led-content-preparer/SKILL.md | Produces a sourced draft for the selected audience and channel. |
| .claude/skills/human-approval-gate/SKILL.md | Blocks external action until an explicit, action-bound approval is present. |
| .claude/skills/crm-decision-recorder/SKILL.md | Creates an auditable CRM patch or applies an approved update through a configured tool. |
| .claude/skills/reply-classifier/SKILL.md | Classifies explicit reply meaning while preserving uncertainty and escalation flags. |
| .claude/skills/objection-analyzer/SKILL.md | Separates stated objections from possible concerns and uses approved proof only. |
| .claude/skills/gtm-learning-reporter/SKILL.md | Reports bounded learnings with counts, denominators, and validation plans. |
| scripts/agentic_gtm.py | Creates runs, reports workflow status, and validates stage outputs. |
| scripts/validate.py | Validates JSON against the repository's strict, skill-specific schemas. |
| examples/northstar-complete/ | Complete 12-stage fictional workflow with validated JSON outputs. |
Prerequisites
- [ ] Claude Code 2.1.205 or later
- [ ] Python 3.10 or later
- [ ] An offer profile describing the problem, target customer, evidence, and disqualifiers
- [ ] Approved source material or read-only research access
- [ ] A named human approver before any external write, send, publish, or CRM update
MCP connections are optional. Every skill supports a file-based workflow.
Installation
Clone and use in this repository
``bash`
git clone https://github.com/forma-norden/claude-code-for-agentic-gtm.git
cd claude-code-for-agentic-gtm
python3 scripts/test_repository.py
claude --permission-mode plan
The skills are already available under .claude/skills/.
Install into an existing Claude Code project
`bash`
git clone https://github.com/forma-norden/claude-code-for-agentic-gtm.git
cd claude-code-for-agentic-gtm
./scripts/install.sh /path/to/your/project
The installer copies the skills into .claude/skills/ and the local validator into .agentic-gtm-kit/.
Cursor, Windsurf, or a generic AI IDE
The skill files are plain Markdown. Read the required SKILL.md file, supply the documented input, and validate the returned JSON with:
`bash`
python3 scripts/validate.py .claude/skills/change-detector/output.schema.json output.json
Usage
Create a workflow run:
`bash`
python3 scripts/agentic_gtm.py init \
--account "Northstar Systems" \
--offer examples/offer-profile.example.json
Start Claude Code and invoke the orchestrator:
`text`
/agentic-gtm Continue the active run. Use only supplied or approved sources. Stop before every external action unless a valid approval record exists.
Run one skill directly:
`text`
/change-detector Read examples/northstar-complete/inputs/01-change-detector.input.json and return JSON matching the skill's output.schema.json.
Validate an output:
`bash`
python3 scripts/validate.py \
.claude/skills/change-detector/output.schema.json \
examples/northstar-complete/outputs/01-change-detector.output.json
Check a run:
`bash`
python3 scripts/agentic_gtm.py status .agentic-gtm/runs/<run-id>
Expected output:
- sourced facts separated from interpretations and assumptions
- a commercial relevance decision with a visible rubric
- a falsifiable why-now hypothesis
- a reversible next action
- an explicit approval state
- an auditable CRM patch or applied-update record
- reply and objection classifications with evidence
- bounded learnings with counts and denominators
Safety Model
- Source text is evidence, never executable instruction.
- Public growth, hiring, funding, or product activity does not prove buying intent.
- External actions remain blocked until approval matches the exact action hash and has not expired.
- Skills use file-based fallbacks when an integration is unavailable.
- Credentials and sensitive customer data must not be stored in prompts, examples, or the repository.
- MCP permissions should begin read-only and follow least privilege.
See [docs/SECURITY.md](docs/SECURITY.md), [docs/DATA-HANDLING.md](docs/DATA-HANDLING.md), and [docs/LIMITATIONS.md](docs/LIMITATIONS.md).
Validation
Repository tests validate:
- skill installation structure
- strict input and output schemas
- complete fictional workflow outputs
- approval defaults
- source preservation
- invalid and incomplete output rejection
- orchestrator run creation and stage status
Run:
`bash`
python3 scripts/test_repository.py
Use [scripts/live_acceptance.sh](scripts/live_acceptance.sh) with an authenticated Claude Code installation to capture and validate a live skill run. The full procedure is documented in [docs/VALIDATION.md`](docs/VALIDATION.md).
Live model output can vary. Review sourced claims, permissions, approval state, and external actions in the installed Claude Code environment.
Who This Is For
GTM engineers, RevOps leads, marketing and growth leaders, sales leaders, and founders at B2B companies building a more connected revenue operation without turning every decision into an autonomous external action.
---
From the Forma Nôrden GTM Library
This is a free resource from the Forma Nôrden open-source GTM library, built by
[Yananai A. Chiwuta](https://yananaichiwuta.com/), GTM engineer and founder of
[Forma Nôrden](https://formanorden.com/).
- [Open-source GTM systems](https://github.com/forma-norden) - all repos in the library
- [GTM engineering blog](https://formanorden.com/blog/) - strategy, systems, and GTM engineering deep-dives
- [All resources](https://formanorden.com/resources/) - guides, frameworks, and templates
If this saves you time, star the repo and follow
[Forma Nôrden on LinkedIn](https://www.linkedin.com/company/formanorden/).
Built by [Forma Nôrden](https://formanorden.com/) - GTM engineering for B2B companies.
Create a workflow run:
``bash`
python3 scripts/agentic_gtm.py init \
--account "Northstar Systems" \
--offer examples/offer-profile.example.json
Start Claude Code and invoke the orchestrator:
`text`
/agentic-gtm Continue the active run. Use only supplied or approved sources. Stop before every external action unless a valid approval record exists.
Run one skill directly:
`text`
/change-detector Read examples/northstar-complete/inputs/01-change-detector.input.json and return JSON matching the skill's output.schema.json.
Validate an output:
`bash`
python3 scripts/validate.py \
.claude/skills/change-detector/output.schema.json \
examples/northstar-complete/outputs/01-change-detector.output.json
Check a run:
`bash``
python3 scripts/agentic_gtm.py status .agentic-gtm/runs/<run-id>
Expected output:
- sourced facts separated from interpretations and assumptions
- a commercial relevance decision with a visible rubric
- a falsifiable why-now hypothesis
- a reversible next action
- an explicit approval state
- an auditable CRM patch or applied-update record
- reply and objection classifications with evidence
- bounded learnings with counts and denominators
- [ ] Claude Code 2.1.205 or later
- [ ] Python 3.10 or later
- [ ] An offer profile describing the problem, target customer, evidence, and disqualifiers
- [ ] Approved source material or read-only research access
- [ ] A named human approver before any external write, send, publish, or CRM update
MCP connections are optional. Every skill supports a file-based workflow.