# The Manifesto for AI-Driven Development

Four values and twelve principles for developers who build software with AI as a deliberate partner.

Version 1.1.1

## Preamble

As AI-Driven Developers, we are discovering better ways of building software by working with AI as a deliberate partner — and helping others do the same.

## Definition

**AI-Driven Development** (AIDD), noun /ˌeɪ.aɪ ˈdrɪv.ən dɪˈvɛl.əp.mənt/

A way of building software in which a developer works with AI as a deliberate partner — planning, decomposing, and reviewing every change — while remaining the architect accountable for what ships.

Related practices: AI-assisted development, AI pair programming, agentic coding, spec-driven development, context engineering.

### AIDD vs vibe coding

| Axis | AIDD | Vibe coding |
| --- | --- | --- |
| Posture | Deliberate partnership | Improvisation — “give in to the vibes” |
| Authorship | You are the architect; every line is yours | The model drives; you forget the code exists |
| Review | Same standards as hand-written — review, test, reject | Accept the output and move on |
| Planning | The first act of collaboration | Prompt and pray |
| Understanding | Extends what you already understand | Ships what you don’t |
| When it stalls | A signal to decompose further | Regenerate until it runs |
| Output | Production software, without debt | Quick experiments, demos, throwaways |
| Best for | Software meant to last | Prototypes you’ll discard |

# Values

## Value 1: Method over Model

<strong>Bet on the method, not the model — every LLM release will fade</strong>.<br> Your method endures and standardizes your team’s practices across models, vendors, and versions.

## Value 2: Ownership over Delegation

<strong>You own what you ship — even what the AI wrote</strong>.<br>Every commit is signed by a human; the AI accelerates, you decide.

## Value 3: Understanding over Acceptance

<strong>Don’t accept what you don’t understand</strong>.<br>The AI is your collaborator — not your replacement.

## Value 4: Outcome over Output

<strong>Writing code is easy. Creating useful outcomes is not</strong>.<br>Ship small, learn quickly, and measure results for users, product reliability, and the business — not lines, tokens, or commits.

# Principles

## Principle 01

Ship in verified increments.

Working software is the output of spec, plan, tests, review, and release.

## Principle 02

Start from a clear spec.

A story, PRD, or spec gives AI work a contract humans can inspect.

## Principle 03

Plan before generation.

The plan is the handoff between human judgment and agent execution.

## Principle 04

Decompose until delegable.

Small tasks survive async work, model changes, and team handoffs.

## Principle 05

Craft context with the care we give code.

Specs, markdown, rules, skills, and memory deserve the same quality bar as software.

## Principle 06

Capture learning for reuse.

Each lesson becomes memory, a rule, a decision, or a skill others can use.

## Principle 07

Own every shipped change.

AI may produce it; humans approve, merge, and maintain it.

## Principle 08

Do not delegate what you cannot evaluate.

Use AI where you can judge tradeoffs, read the output, and prove behavior.

## Principle 09

Improve the system before forcing the model.

When output degrades, clarify boundaries, names, tests, and context before asking harder.

## Principle 10

Rebuild context when coherence drops.

Reset from memory, docs, rules, and the current plan instead of improvising.

## Principle 11

Treat AI failure as a signal.

When the model stalls, inspect scope, context, assumptions, and tests before pushing harder.

## Principle 12

Make the practice transferable.

A shared method lets different people produce the same quality of code across tools.

## Sign

Signing is a public act recorded in the source repository. One YAML file, one pull request, no backend.

Repository: https://github.com/ai-driven-dev/manifest
Canonical URL: https://www.ai-driven-development.org/