# IntentForge: Deliberate Prompts and Bounded AI Agents

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

IntentForge (https://intentforge.ai.studio) turns rough prompts and agent loops into deliberate, bounded specifications, then proves the difference with live A/B runs on the model API across 12 frontier models.

## Live Benchmark Proof (Computed from Live API Runs)
- **44.6% median tokens saved on agent loops** · **37.3% on prompts** · measured on **11 live benchmarks**.
- **Order & Return Resolution Agent**: 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%**.

## What a Deliberate Spec Contains
- **Strict output contract**: every run returns the same shape.
- **Tool call contracts**: preconditions and order, with no duplicate calls.
- **Explicit conditions**: thresholds, edge cases and escalation rules written down, not left to the model.
- **Hard step ceiling**: a fixed limit, then a terminal state. No runaway loops.

## Public Pages & Essays
- [Home](https://intentforge.ai.studio/): Deliberate prompts and bounded AI agents.
- [Benchmarks](https://intentforge.ai.studio/benchmarks): Live A/B runs comparing rough prompts with compiled specs.
- [Studio](https://intentforge.ai.studio/studio): Compile prompts and agent loops and test live against your original.
- [Glossary](https://intentforge.ai.studio/glossary): Formulas, rubrics and effective tokens per usable result.
- [Essays & Guides](https://intentforge.ai.studio/blog): Five essays on deliberate prompting, cost per usable result, stopping agent loops, AGENTS.md, and OWASP guardrails.