A context engineering framework that makes AI write like you, not like AI.
AI-generated content starts strong but flattens into generic corporate filler within minutes. Every chat session rebuilds context from scratch — losing your voice, inventing metrics, and making claims your business can't back. Teams end up spending more time editing AI output than they saved generating it.
The Context Engineering Kit replaces ad-hoc prompting with persistent truth files — structured JSON that tells the AI who you are, who you serve, what you offer, and what you're allowed to say. Every content request loads these files automatically, so the AI never starts from zero.
Four Pillars of Truth:
| Pillar | File | What It Controls |
|---|---|---|
| Voice DNA | voice-dna.json |
Tone, cadence, phrasing, boundaries |
| Ideal Client Profile | icp.json |
Pains, desires, objections, language |
| Business Profile | business-profile.json |
Offers, positioning, pricing logic |
| Claims Policy | claims-policy.json |
What can and cannot be stated |
Format-Specific Skills layer on top — one markdown file per content type (landing page, cold email, LinkedIn post, sales deck) with structure, inputs, and QA rules baked in. Voice and claims never live in skills; they come from truth files. This separation is what prevents drift.
Hierarchical Control eliminates ambiguity when rules conflict:
Truth & Claims Policy → always wins
Voice DNA → wins on tone
ICP → wins on language
Business Profile → wins on offers
Skill → wins on format
- Structured JSON Context — machine-readable truth files, not prose prompts
- Skill Architecture — one file per content type with inputs, structure, QA checklist, and constraints
- Claims Policy Engine — blocks hallucinated metrics, fabricated testimonials, and unsupported guarantees at the system level
- Static Explorer UI — vanilla HTML/CSS/JS dashboard with markdown rendering, trust-level badges, and file metadata — zero build tooling
- Tool-Agnostic — works with Claude, ChatGPT, Cursor, Windsurf, or any RAG-capable system
- Knowledge Compounding — gold-standard examples and reusable templates feed back into the system, so your best work fuels the next asset
- Node.js >= 18
- An AI tool that supports system prompts (Claude, ChatGPT, Cursor, Windsurf)
git clone https://github.com/RCushmaniii/context-writing-system.git
cd context-writing-system
npx serve . -p 3000Run the creator prompts in your AI tool to generate your truth files:
prompts/voice-dna-creator.mdgeneratescontext/core/voice-dna.jsonprompts/icp-creator.mdgeneratescontext/core/icp.jsonprompts/business-profile-creator.mdgeneratescontext/core/business-profile.json
Then point your AI tool to CLAUDE.md and start writing.
├── context/core/ # Source-of-truth JSON profiles
├── skills/ # Content-type expertise (one file per format)
├── knowledge/ # Examples, templates, references
│ ├── content/ # Gold-standard published examples
│ ├── templates/ # Reusable blocks (CTAs, proof, bios)
│ ├── drafts/ # Work-in-progress
│ └── notes/ # Raw ideas and research
├── prompts/ # Creator interview prompts
├── survey/ # Discovery survey definitions
├── docs/ # Supporting documentation
├── archive/ # Versioned history
└── CLAUDE.md # System contract
| Metric | Before | After |
|---|---|---|
| Context setup | Rebuilt every session | Loaded once from truth files |
| Draft quality | Generic, needs heavy editing | On-brand, ready for review |
| Editing time | ~45 min per asset | ~5 min per asset |
| Hallucinated claims | Frequent and unchecked | Blocked by claims policy |
| Knowledge reuse | Scattered across chat tabs | Compounds with every asset |
| Tool | Status |
|---|---|
| Claude Projects / Claude Code | Fully supported |
| Cursor | Fully supported |
| Windsurf | Fully supported |
| ChatGPT | Fully supported |
| Any RAG-capable system | Compatible |
Markdown + JSON. No vendor lock-in.
Robert Cushman — Business Solution Architect & Full-Stack Developer Guadalajara, Mexico
info@cushlabs.ai | GitHub | LinkedIn | Portfolio
Proprietary — All rights reserved. See LICENSE for details.
