Overnight AI Development

Tests as specifications

"Build a user login system" is ambiguous. An AI will guess what you want, and it'll guess wrong.

But tests are precise. A test says exactly what should happen: this input produces that output, this action triggers that event, this error shows that message. No ambiguity.

So the workflow becomes: analyst writes tests first, human approves the test specifications, then the dev agent implements to pass those tests. The expensive thinking happens upfront in the analysis phase. Implementation becomes mechanical.

File-based kanban

The whole thing runs on files moving through directories:

project/
├── backlog/           # human adds feature requests
├── analysis/          # AI analyst writes tests
├── approved/          # human approves test specs
├── development/       # AI dev implements
└── done/              # completed and validated

Everything visible, debuggable, version-controlled. No magic databases, no hidden state. Just files.

The overnight workflow

Before bed: review analyses, approve what looks good, reject what doesn't.

Overnight: kick off the development agent on low priority, trickle power. It processes approved items, runs tests, commits results.

Morning: check what got built. Run integration tests. Move good implementations to done, send problems back to analysis with notes.

The model swaps in and out to save memory. Rate limiting keeps power usage reasonable. Everything checkpointed so you can rollback if something goes wrong.

Why this works

Compute efficiency: Analysis is cheap. Human review is the gate before expensive development. Most bad ideas get killed early.

Quality: Tests written first means clear acceptance criteria. No "is this what you wanted?" conversations.

Rollback: Git checkpoint before overnight run, commit after. Something went wrong? Reset to stable tag.

Agile without meetings: No sprint planning, no standups, no estimation. Just a backlog and tests.

The goal

A self-managing development team that processes your backlog while you sleep. Tested implementations ready for review in the morning. Runs on a raspberry pi with a solar panel if you want.

You stay the product owner - making decisions about what to build. The AI handles the mechanical work of how to build it, following your approved specifications.

Files and scripts doing real work.