An IDE for Persuasion
Treat psychological influence as executable code.
Write text that compiles into specific human reactions.
100% offline. Runs entirely in your browser. No server. No API keys. Your data never leaves your device.
Problem • Concept • How It Works • Modules • Tech Stack • Getting Started • Roadmap
Every day, billions of messages are sent — cold emails, job applications, sales pitches, landing pages. Most of them fail. Not because the information is wrong, but because the psychology is wrong.
| What you wrote | Why it fails | What works |
|---|---|---|
| "I hope to hear from you" | Hedging → signals low status | "Let's align on next steps Thursday" |
| "Was responsible for managing a team" | Passive voice → weak authority | "Led a 12-person engineering team" |
| "Buy now!" | High friction word → triggers resistance | "Unlock instant access" |
| "Our product is the best" | Unquantified claim → zero credibility | "Trusted by 4,200+ teams since 2021" |
These aren't style preferences. They're predictable psychological patterns backed by decades of behavioral science research.
text-as-code treats these patterns as rules — like a linter for code, but for human persuasion.
What if text had a compiler?
INPUT: "I think our product might be able to help you."
^^^^ ^^^^^
WARN: Hedging WARN: Uncertainty marker
OUTPUT: "Our platform eliminates 6 hours of manual work per week."
✅ Authority ✅ Quantified impact
text-as-code is a real-time analysis engine that:
- Parses your text against psychological rule sets
- Highlights weak spots (🔴), missed opportunities (🟡), and strong patterns (🟢)
- Explains the cognitive science behind each flag
- Suggests 1-click rewrites — powered by an AI running entirely inside your browser
Think ESLint for persuasion. Or Grammarly, but for influence instead of grammar.
Every piece of text runs through two complementary engines — both running 100% client-side:
A curated library of 50+ regex/NLP patterns that catch common psychological missteps in milliseconds:
{
"id": "WEAK_HEDGE",
"pattern": "\\b(I think|maybe|perhaps|I hope)\\b",
"severity": "warning",
"trigger": "cognitive_weakness",
"psychology": "Hedging activates uncertainty bias in the reader, reducing perceived competence by ~40%.",
"suggestion": "Remove the hedge. State with authority."
}Rules are tagged by module — a resume gets different checks than a sales email.
For nuances that patterns can't catch — tone, framing, narrative arc — an LLM runs directly in your browser via WebGPU:
"Your cold email uses reciprocity correctly in paragraph 1,
but the CTA in paragraph 3 creates cognitive overload by
presenting 3 choices. Reduce to a single, binary CTA."
No API calls. No server. The model runs on your GPU locally.
| Runtime | Technology | When it's used |
|---|---|---|
| WebLLM | WebGPU-accelerated (30-70 tok/s) | Modern browsers with GPU |
| Transformers.js | WASM fallback (4x faster with WebGPU in v4) | Older browsers, no GPU |
Models are downloaded once and cached in IndexedDB for fully offline use.
The editor feels like VS Code meets Grammarly:
| Color | Meaning | Example |
|---|---|---|
| 🔴 Red | High cognitive friction — reader will bounce | Sentence over 40 words |
| 🟡 Yellow | Weak framing — missed persuasion opportunity | Passive voice, no CTA |
| 🟢 Green | Strong — psychologically effective | Quantified claim, pattern interrupt |
Hover over any highlight → see the psychological principle, the research behind it, and a 1-click rewrite.
A Persuasion Score (0-100) gives an at-a-glance quality metric, with a radar chart showing coverage across psychological triggers.
The system routes text through discipline-specific psychological rule sets:
| Primary Triggers | What the engine does |
|---|---|
| Loss Aversion | Flags missing scarcity/urgency framing |
| Risk Reversal | Detects absence of guarantees or social proof |
| Assumed Close | Rewrites weak asks into confident next-step language |
| Primary Triggers | What the engine does |
|---|---|
| Authority & Competence | Strips passive voice, injects power verbs |
| Halo Effect | Ensures strong opening creates positive bias cascade |
| Quantified Impact | Forces "increased sales" → "drove 20% revenue increase" |
| Primary Triggers | What the engine does |
|---|---|
| Curiosity / Info Gap | Optimizes subject lines to force opens |
| Pattern Interrupt | Flags generic openings that get deleted |
| Cognitive Ease | Measures reading time, flags complexity |
Zero backend. Everything runs in the browser.
| Layer | Technology | Why |
|---|---|---|
| Build | Vite | Fast HMR, small bundles |
| UI | React 18 | Component model, hooks, concurrent features |
| Editor | TipTap (ProseMirror) | Custom decorations for highlights |
| AI (Primary) | WebLLM (@mlc-ai/web-llm) |
OpenAI-compatible API, WebGPU, 30-70 tok/s |
| AI (Fallback) | Transformers.js v4 (@huggingface/transformers) |
WASM + WebGPU, 1200+ models, offline caching |
| State | Zustand | Lightweight, no boilerplate |
| Storage | IndexedDB | Persist docs & cached models offline |
| Styling | Vanilla CSS | Full control, no framework overhead |
| Model | Size | Best for |
|---|---|---|
| Phi-3.5-mini-instruct (q4) | ~2 GB | Deep analysis, best quality |
| SmolLM2-1.7B-Instruct (q4) | ~1 GB | Fast analysis, lower-end GPUs |
| Qwen2.5-1.5B-Instruct (q4) | ~1 GB | Transformers.js fallback |
text-as-code/
├── src/
│ ├── main.jsx
│ ├── App.jsx
│ ├── index.css # Design system
│ │
│ ├── engine/ # 🧠 Psychological analysis core
│ │ ├── rules/ # Static rule definitions (JSON)
│ │ │ ├── common.json # 15 cross-module rules
│ │ │ ├── sales.json # 7 sales-specific rules
│ │ │ ├── resume.json # 7 resume-specific rules
│ │ │ └── cold-email.json # 8 cold email rules
│ │ ├── modules.js # Module registry + trigger weights
│ │ ├── static-analyzer.js # Pass 1: regex rule matcher
│ │ ├── llm-analyzer.js # Pass 2: in-browser LLM (Phase 4)
│ │ ├── prompts.js # Module-specific system prompts
│ │ └── scorer.js # Persuasion score (0-100)
│ │
│ ├── ai/ # 🤖 In-browser AI runtime (Phase 4)
│ │ ├── provider.jsx # React context for AI
│ │ ├── webllm.js # WebLLM adapter
│ │ ├── transformers.js # Transformers.js adapter
│ │ └── model-manager.js # Download, cache, fallback
│ │
│ ├── components/ # 🎨 UI
│ │ ├── Editor/
│ │ │ └── Editor.jsx # TipTap editor + highlight plugin
│ │ ├── Suggestions/
│ │ │ └── SuggestionList.jsx # Issue cards with psychology + fixes
│ │ ├── Score/
│ │ │ └── ScorePanel.jsx # Ring gauge + trigger bars
│ │ └── ModuleSwitcher.jsx # Sales / Resume / Cold Email
│ │
│ └── store/
│ └── editor-store.js # Zustand state management
│
├── index.html
├── package.json
├── vite.config.js
└── README.md
- A modern browser with WebGPU support (Chrome 113+, Edge 113+, Safari 18+)
- Node.js ≥ 18 (for development only)
- No API keys needed. No server needed.
# Clone
git clone https://github.com/AdrianBesleaga/text-as-code.git
cd text-as-code
# Install
npm install
# Start dev server
npm run devOpen http://localhost:5173 — paste your text, select a module, and watch the psychological analysis in real time.
- The app loads instantly with static rules — no download needed
- Click "Download AI Model" to enable deep LLM analysis (~1-2 GB, one-time)
- Once cached, the app works fully offline — disconnect your network and try it
Your text never leaves your browser. There is no server. There are no API calls. The AI model runs on your local GPU via WebGPU. All data is stored in IndexedDB on your device. You can verify this by opening DevTools → Network tab while using the app.
This makes text-as-code safe to use with:
- Confidential business proposals
- Salary negotiation emails
- Legal correspondence
- Anything you wouldn't paste into ChatGPT
- 37 psychological rules across 3 modules
- TipTap editor with color-coded highlights (red/yellow/blue)
- Persuasion Score (0-100) with trigger coverage bars
- Inline issue cards with psychology explanation + fix suggestion
- Module switching (Sales, Resume, Cold Email)
- WebLLM integration (WebGPU)
- Transformers.js fallback (WASM)
- Model download manager with progress UI + model selection
- Deep contextual analysis via local LLM
- Document persistence (IndexedDB)
- Export to PDF / Markdown / Clipboard
- Custom rule editor (create your own psychological rules)
- Side-by-side before/after comparison view
- Chrome Extension (inject into Gmail, LinkedIn, Google Docs)
- PWA / installable offline app
- Community rule packs (share & import modules)
Text is the most underrated technology. It's the interface between every human interaction — hiring, selling, persuading, connecting. Yet we treat writing as an art, not a science.
text-as-code treats writing as engineering:
- Patterns are bugs. They can be detected and fixed.
- Principles are functions. They can be composed and applied.
- Outcomes are tests. They can be measured and optimized.
Write text that compiles into the reaction you want.
MIT
text-as-code — Because every word is an instruction to the human mind.