Specification · v0.1.1 draft

AI Development Mode

A self-declared convention for communicating how human developers and generative AI systems shared agency during software development.

Core rule: classify the typical workflow for substantive changes. The modes describe process and agency, not quality, trust, skill, or AI-generated line count.

The modes

Autonomous

After receiving an initial goal, prompt, or specification, AI performs substantially the complete development process with little or no intermediate human direction.

Supervised

AI independently owns substantial multi-step development work, while a human reviews meaningful milestones and can redirect the work.

Collaborative

Human and AI develop the software together through frequent, iterative interaction, with both materially influencing decisions and implementation.

Directed

The human determines requirements, architecture, and implementation direction; AI is delegated bounded development tasks inside that direction.

Assistive

The human performs the core development work, while AI provides supporting assistance such as debugging, tests, documentation, explanations, review, or small snippets.

Incidental

Generative AI use is occasional, local, and not materially involved in technical direction or the substantive implementation workflow.

None

No generative AI was intentionally used in the declared development scope.

Mixed

Two or more materially different workflows are common and no single headline mode honestly characterizes the declared scope. Mixed requires a profile or explanation.

The modes above are ordered from greatest to least AI autonomy. The order is used for classification only and says nothing about quality, maturity, or trust.

Choosing a mode

If two adjacent modes fit equally well, choose the mode reflecting greater AI autonomy and explain the nuance in the profile. If non-adjacent workflows are both common, do not average them: use Mixed or separate scopes.

Optional profile

Field	Values
planning, architecture, implementation, testing, review	ai-led · shared · human-led · not-applicable
human_review_frequency	final · milestone · frequent · continuous · not-applicable
autonomous_scope	project · feature · task · suggestion · none

Machine-readable form

The conventional filename is .aidm.json. A declaration must contain spec_version, mode, and scope.

{
  "spec_version": "0.1.1",
  "mode": "collaborative",
  "scope": { "type": "repository" }
}

Badge

The recommended label is AI Development: <Mode>. Colors should be neutral and consistent; do not use a red-to-green scale that suggests a rating.

Non-goals

AIDM does not establish copyright authorship, certify security/quality, detect AI-written code, quantify developer skill, or prescribe whether AI should be used.

Full normative source

The repository's SPECIFICATION.md is the normative source for the draft, including conformance and versioning details.