# IntentForge — Complete Reference: Deliberate Prompts & Bounded AI Agents

Canonical URL: https://intentforge.ai.studio
Core message: People speak in gists. Models need it spelled out.

## 1. The Thesis: Software Used to Make Us Deliberate. Chat Stopped.
Forms wouldn't submit until every field was filled. Dropdowns allowed only valid choices. Checkout flows turned a vague wish into explicit decisions. For fifty years, interfaces did the deliberating for us. Chat removed all of it: now anyone can type anything, and the model fills every gap by guessing. You pay for each guess in output tokens, wrong tool calls and retries. IntentForge puts that layer back, without making anyone fill in a form.

## 2. How It Works
1. **Describe it roughly**: A prompt or an agent goal, its tools, a step budget and the target model.
2. **Compile**: IntentForge makes the intent explicit: schema, constraints, tool contracts, stop rules.
3. **Run a live A/B trial**: Your original and the compiled spec both run on the real API, with exact token counts recorded.
4. **Check against a rubric**: Task checks plus OWASP, NIST AI RMF and BFCL presets.
5. **Export**: Straight into your stack, in nine formats (JSON Schema, Structured Markdown, Bullet points, Natural prose, Code / SQL, Agent manifest, AGENTS.md, MCP tool and loop JSON, A2A handoff).

## 3. Live Benchmark Proof
- Measured across **11** live-tested benchmarks and **12** frontier models.
- **44.6%** median tokens saved on agent loops · **37.3%** median tokens saved on prompts.
- On our **Order & Return Resolution Agent**, the compiled spec was only **8%** shorter per run (3,032 → 2,790 tokens). But first-try pass went from **50%** to **100%**, so tokens per usable result fell **54%**.

## 4. Frequently Asked Questions
### Isn't this just prompt shortening?
No. Deliberate specs are often longer than the original. They cost less per usable result because the model stops guessing, fails less and retries less.

### How are the savings measured?
Every compile runs a live A/B trial: your original and the compiled spec both run on the model API, and IntentForge records the exact input and output token counts and the pass rate.

### Which models does it support?
12 frontier models from Google, OpenAI, Anthropic, Meta and DeepSeek, with costs priced from each provider's published rates.

### Won't models just get better at guessing what I mean?
The gaps are in what you didn't say, not in the model. No model can infer a refund limit you never stated. Stating it once in the spec is cheaper than the model asking, or guessing, on every run.

### Do you store my prompts?
Your prompts and specs stay private in your workspace. If you opt in to benchmark telemetry, only numeric token and retry metrics are shared. You can export or delete everything at any time.