Loop Engineering

Stop asking coding agents to remember your entire project. Build a persistent source of truth they are forced to follow.

Idea Discussion Specification Implementation Test Inspect Correct Verify Repeat

How it works

Loop engineering is not about writing a larger prompt.

Instead of: Prompt Code

use: Specification Implementation Test Failure Correction Verification Repeat

The model can change. Claude Code, Codex, Antigravity, or another coding agent can execute the workflow.

The source of truth stays outside the model.

Project structure

loop-engineering/
│
├── 00_SOURCE_OF_TRUTH.md
├── 01_PRD.md
├── 02_ARCHITECTURE.md
├── 03_SPECIFICATION.md
├── 04_TASKS.md
├── 05_AGENT_RULES.md
├── 06_VERIFICATION.md
├── 07_DECISIONS.md
├── 08_PROGRESS.md
│
└── prompts/
    ├── 01_DISCOVERY_PROMPT.md
    ├── 02_GENERATE_FILES_PROMPT.md
    ├── 03_IMPLEMENTATION_PROMPT.md
    ├── 04_TASK_EXECUTION_PROMPT.md
    ├── 05_VERIFICATION_PROMPT.md
    └── 06_FINAL_AUDIT_PROMPT.md

Files

• 00_SOURCE_OF_TRUTH.md: The master specification that agents read first.
• 01_PRD.md: Product Requirements Document.
• 02_ARCHITECTURE.md: System architecture rules.
• 03_SPECIFICATION.md: Technical specification.
• 04_TASKS.md: Granular execution checklist.
• 05_AGENT_RULES.md: Hard rules the agent must not break.
• 06_VERIFICATION.md: Acceptance criteria and tests.
• 07_DECISIONS.md: Log of architectural decisions (ADRs).
• 08_PROGRESS.md: Living document of completed features.

Prompts

• 01 Discovery: Used for brainstorming and requirement gathering.
• 02 Generate Project Files: Used to generate the initial loop-engineering folder.
• 03 Start Coding Agent: The initial prompt to bootstrap the coding agent.
• 04 Execute One Task: Used for single, focused task execution.
• 05 Verification Loop: Used to verify completed code against requirements.
• 06 Final Audit: Final review before declaring a task complete.

On OpenDeck the six prompt bodies are not filled in yet; each still shows a "paste the prompt here" placeholder.

Start here

1. Take an idea.
2. Use the Discovery prompt with Claude or another reasoning model.
3. Discuss and challenge the idea before coding.
4. Generate the project-control folder.
5. Place that folder inside the coding repository.
6. Give the coding agent the Start Coding Agent prompt.
7. Let the agent implement one task at a time.
8. Verify actual results.
9. Feed failures back into the implementation loop.
10. Finish only when every acceptance criterion has evidence.

The downloadable kit is marked "coming soon" on OpenDeck.