sisig.ai launches agent-doctrine for AI code generation

Getting an agent to write code is a solved problem, but ensuring it writes the right code, proves it functions properly, and hands over a pull request you would actually merge remains a significant challenge.

Over the past year, the team at sisig.ai has iterated on a collection of prompts, role definitions, and skills to bridge this gap, recently open-sourcing the result as agent-doctrine under an MIT license.

Every AI agent operating within this framework follows a strict sequence of eight distinct steps:

• scout
• gate
• plan
• implement
• review
• consolidate
• verify
• ship

The agent you interact with acts as the lead, responsible for planning and delegation while avoiding manual code writing unless explicitly instructed. A major shift in this workflow is that prior to any edits, the lead must answer fifteen questions from TENETS.md in the transcript, complete with verifiable evidence. If the agent fails to answer these criteria, it is not permitted to write code.

Stock photo for illustration only, not from the actual event

Furthermore, handoffs utilize files with a fixed shape called HANDOFFS.md, allowing a specification drafted in Claude Code to be seamlessly picked up by Codex or Cursor.