Implement

Contents

• Overview
• Running an Implement Job
  • Web app
  • CLI
• Choosing a Backend
• Max Turns
• Output
• Tips

---

Overview

Implement takes a compose or transmute output file and drives an AI agent to write a complete, working implementation from the spec. The agent reads the spec, plans the implementation, and writes code files directly to an output directory.

---

Running an Implement Job

Web app

After a Compose or Transmute job completes, the job detail page offers an Implement button that pre-fills the spec path and launches the Implement workflow.

You can also navigate to the Projects or Job History screen and start an Implement run from any completed compose job.

CLI

# Implement from a compose output file
otx implement Output/Composition/my-design/compose-output.md \
  --project my-new-system \
  --output ./generated/my-new-system

# With explicit backend and model
otx implement path/to/spec.md \
  --orchestrator OpenAI \
  --api-key sk-... \
  --model gpt-4o \
  --max-turns 300 \
  --timeout 90

See CLI: implement for the full option reference.

---

Choosing a Backend

For the Implement step, ClaudeCode is strongly recommended when available. The claude CLI agent uses its own file-reading and file-writing tools to explore the output directory and write code directly — no file packing or intermediate steps required. It handles large, multi-file implementations naturally.

For OpenAI or Ollama backends, the agent is driven via the FileSystemPlugin, which provides read/write tools over the output directory.

---

Max Turns

Implement runs are longer than Compose runs. The default --max-turns is 200, which is appropriate for small-to-medium specifications. For complex specs generating many files, increase it:

otx implement spec.md --max-turns 400 --timeout 120

---

Output

Generated code is written to the --output directory (default: Output/Implementation/<project>/). The agent creates all files and directories needed for the implementation.

The output directory is not cleared before a run. If you re-run Implement on the same directory, the agent will see its previous output and can build on or correct it.

---

Tips

• Provide a high-quality spec. The quality of the implementation is directly proportional to the quality of the compose/transmute output. If the spec is vague, the implementation will be too. Consider adding a user ethos to set concrete standards.
• Use the thick model. The Implement step defaults to the thick model (the most capable one). Don’t downgrade unless you have a cost reason — complex multi-file implementations benefit significantly from the largest context window and strongest reasoning.
• Check the output directory first. Implement does not clean the output directory. If you’re re-running after a partial attempt, review what’s already there so you understand what the agent will see.