messaging-gpt-builder

A Claude Code skill system that helps Product Marketing Managers build their own MessagingGPT — a complete AI-powered messaging intelligence assistant trained on YOUR company's knowledge.

What This Is

You can't copy someone else's MessagingGPT. It's trained on their company's confidential data — their ICP, competitive intel, customer transcripts, proof points. But you CAN replicate the architecture and reasoning approach that makes it work.

This project is the meta-system. It guides you through building your own messaging knowledge base and instruction set, then packages it for deployment as a Custom GPT or Claude Code project.

What You Get

10 skills that chain together:

EXTRACT                    BUILD (per file)              VALIDATE
┌──────────┐              ┌──────────────┐              ┌───────────┐
│/interview│──extracts──▶│ /build-file  │──drafts────▶│  /audit   │
│ /ingest  │              │              │              │           │
└──────────┘              └──────────────┘              └─────┬─────┘
                                                              │
                          PERSONALIZE                    ┌────▼──────┐
                          ┌──────────┐                   │ /assemble │
                          │/brain-   │                   │           │
                          │ build    │                   └─────┬─────┘
                          │/brain-   │                         │
                          │ update   │                   ┌────▼──────┐
                          └──────────┘                   │/test-drive│
                                                         └───────────┘

10 scaffolds — reasoning guides (not templates) for each knowledge file:

• Product Knowledge, ICP & Pains, Positioning & Narrative, Differentiation, Allowed Claims, Messaging Matrix, Evidence Vault, Objection Handling, Vertical Playbooks, System Instructions

Universal PMM Brain — 35+ mental models from Dunford, Raskin, Kramer, Walker, Balfour, Verna, Pierri + 7 argumentation patterns + 7 anti-patterns

Dual output — packages for both Custom GPT upload and Claude Code deployment

How It Works

1. Extract Your Knowledge

Feed the system your raw materials — sales transcripts, case studies, competitor pages, product docs. Or let it interview you with Socratic questions that surface knowledge you haven't articulated.

2. Build Your Knowledge Files

The system generates each file using scaffolds as reasoning guides. Not fill-in-the-blank — genuine strategic thinking with mental models applied. Every claim gets a source tag.

3. Quality Gates at Every Step

Director Checkpoint forces hard questions before building. 6-check audit validates each file. Assembly validation catches cross-file contradictions. Test-drive simulates the GPT to find gaps.

4. Deploy

Export as a Custom GPT (markdown files + upload guide) or a Claude Code project (CLAUDE.md + skills + knowledge files).

The Build Sequence

Files are built in dependency order:

Order	File	Key Question
1	Product Knowledge	What does it do?
2	ICP & Pains	Who is it for? What hurts?
3	Positioning & Narrative	How do you frame it?
4	Differentiation	How do you win?
5	Allowed Claims	What can you prove?
6	Messaging Matrix	Full messaging scaffold
7	Evidence Vault	Proof by pain
8	Objection Handling	Resistance patterns
9	Vertical Playbooks	Per-vertical overlays
10	System Instructions	GPT orchestration (built LAST)

Key Design Principles

Reasoning over templates. Scaffolds teach thinking, not form-filling. The ICP scaffold doesn't ask "what is your target market?" — it asks "who should your sales team STOP talking to?"

Source attribution on every claim. If it's not sourced, the GPT shouldn't say it. [source: inferred] is allowed but flagged for enrichment.

Human-in-the-loop. Positioning and messaging are strategic choices, not single answers. The system presents 2-3 angles with trade-offs. The PMM decides.

Editorial guardrails. 50+ banned words, 7 banned sentence patterns, voice standards. The GPT should sound like your company, not like AI.

Self-improving. The observe/inspect-amend loop tracks quality signals and proposes improvements to scaffolds, gates, and brain models based on real usage.

Getting Started

1. Clone this repo into a Claude Code workspace
2. Run /interview to start a discovery session — or place materials in workspace/raw-materials/ and run /ingest
3. Follow the build sequence: /build-file 01, then /build-file 04, etc.
4. Run /audit after each file
5. Run /assemble to package for GPT or Claude Code
6. Run /test-drive to simulate and find gaps

What You Need (Minimum)

Material	Minimum	Ideal
Sales/demo transcripts	5	15-20
Case studies with metrics	3	8-10
Product documentation	1 doc	Feature docs + pricing + integrations
Competitor pages	2 competitors	5 competitors + alternatives

Project Structure

messaging-gpt-builder/
├── CLAUDE.md                    # Orchestrator
├── skills/                      # 10 skills + data flow
├── scaffolds/                   # 10 reasoning guides + principles
├── brain/                       # Universal PMM brain + user brain
├── references/                  # Build sequence, quality gates, editorial guardrails, etc.
├── examples/                    # 3 anonymized reference files
├── workspace/                   # Your working files (raw materials, extracts, drafts)
├── output/                      # Deployable packages (GPT + Claude Code)
├── _logs/                       # Observation logs
└── _versions/                   # Amendment backups

Credits

Built on frameworks from April Dunford, Andy Raskin, Emily Kramer (MKT1), Chris Walker, Brian Balfour, Elena Verna, Sangram Vajre, Anthony Pierri, and Play Bigger.

Architecture informed by real-world messagingGPT deployments.

License

MIT