loop-engineering

Files (53)

• AGENTS.md
• CLAUDE.md
• CONTEXT.md
• LOOP-ENGINEERING-CONCEPT.md
• README.md
• SKILL.md
• SOURCES.md
• CONTEXT.md · _shared/
• collins-level.md · _shared/doctrine/
• commercial.md · _shared/doctrine/
• heart-and-soul.md · _shared/doctrine/
• mobile-mission.md · _shared/doctrine/
• engine-router.md · _shared/routing/
• skill-router.md · _shared/routing/
• example-scorecard.json · _shared/schemas/
• gate.yaml · _shared/schemas/
• project-lock.yaml · _shared/schemas/
• receipt.yaml · _shared/schemas/
• task-node.yaml · _shared/schemas/
• bar-and-gauntlet.md · _shared/standards/
• commercial.md · _shared/standards/
• evidence.md · _shared/standards/
• graph-rules.md · _shared/standards/
• mobile.md · _shared/standards/
• release.md · _shared/standards/
• APPROVAL.md · _templates/RUN/
• BAR.md · _templates/RUN/
• BASELINE.md · _templates/RUN/
• GATES.md · _templates/RUN/
• GAUNTLET.md · _templates/RUN/
• GRAPH.md · _templates/RUN/
• LEARNINGS.md · _templates/RUN/
• PROJECT-LOCK.md · _templates/RUN/
• RECEIPT.md · _templates/RUN/
• ROLLBACK.md · _templates/RUN/
• SPEC.md · _templates/RUN/
• STATE.md · _templates/RUN/
• CROSS-AGENT.md · adapters/
• HERMES-HANDOFF.md · adapters/
• init-run.py · scripts/
• score-gauntlet.py · scripts/
• validate-workspace.py · scripts/
• CONTEXT.md · stages/00_intake/
• CONTEXT.md · stages/01_discovery/
• CONTEXT.md · stages/02_architecture/
• CONTEXT.md · stages/03_graph/
• CONTEXT.md · stages/04_spec/
• CONTEXT.md · stages/05_slice/
• CONTEXT.md · stages/06_build/
• CONTEXT.md · stages/07_verify/
• CONTEXT.md · stages/08_gauntlet/
• CONTEXT.md · stages/09_release/
• CONTEXT.md · stages/10_learn/

SKILL.md (first lines)

# Loop Engineering

One skill. Small context. Full lifecycle.

This file is a router, not the playbook. Do not load the whole workspace.

## Start

1. Read `AGENTS.md`.
2. Read root `CONTEXT.md`.
3. Determine `MODE`: greenfield or brownfield.
4. Inspect before asking questions when a repo/product already exists.
5. Establish a **BAR** before substantial work. If the user did not name one, propose 2-3 specific, fetchable, comparable bars and ask them to choose.
6. Create or resume one folder under `runs/` from `_templates/RUN/`.
7. Follow the next stage named in that run's `STATE.md`.

## Core operating law

`INTENT -> BAR -> LOCK -> EVIDENCE -> GRAPH -> SPEC -> SLICE -> BUILD -> VERIFY -> GAUNTLET -> RELEASE -> LEARN`

The outer loop repeats only when evidence says the current result has not cleared the bar. The task graph exists inside the loop and parallelizes only genuinely independent work.

## Never

- claim completion from code presence, CI, or deployment alone;
- let a builder approve its own work;
- start a rewrite before brownfield inspection;
- run every skill on every task;
- add a framework because it is interesting;
- expose secrets or move owner authority into the browser;
- ship without rollback and production evidence;
- create more active work merely because agents are available.

## End-state words

Use only: `NOT READY`, `READY FOR PREVIEW`, `PREVIEW VERIFIED`, `PRODUCTION VERIFIED`.

`PRODUCTION VERIFIED` requires runtime evidence from the exact released revision.

From executiveusa/vibe-engineering, MIT.

More from executiveusa/vibe-engineering