Apo

Apo

**A proof of concept for intent-driven communication with LLMs.**

Agentic coding involves communicating with an LLM in natural language, leading to conversational drift. Increasingly, this drift also involves turning agency in ideation and decisionmaking over to the coding agents.

Apo explores a simple idea: what if you structured your intent *before* the agent started working in the codebase? Drawing from speech act theory, apo uses six "intention primitives" (WANT, DON'T, LIKE, FOR, ENSURE, TRUST) that capture the dimensions people need to specify so an LLM can capture their intent.

**This is a prototyping tool, not a production workflow system.** The apo tool here is currently a proof of concept, designed for solo developers and quick builds — the kind of work where you're talking to Claude Code and want to get something right on the first pass. It's not (yet) built for engineering teams, CI/CD pipelines, or enterprise workflows. The contribution is the framework (the six primitives and how they interact).

Before / After

**Without structured intent**, a typical prompt:

"Build me an expense tracker with categories"

The agent fills in the blanks with its own preferences. It picks a database, an auth system, an API shape. You get something that works, but the agent made the design decisions, not you.

**With Apo**, you think through the full picture first:

## WANT
- CRUD operations for expenses
- Summary with totals by category

## DON'T
- No external database
- No authentication system

## LIKE
- Stripe's API design

## FOR
- Personal finance prototype
- Single user, runs locally

## ENSURE
- Negative amounts throw an error
- Categories are case-insensitive

## TRUST
- [autonomous] Choose the data structure
- [ask] Modifying the public API signature

Now you've made the decisions, not the agent. Scope, boundaries, context, acceptance criteria, and delegation rules are all explicit before a single line of code gets written.

G