A collection of reusable skills, engineering-discipline , has been released to give AI coding agents structured planning, verification, and cleanup discipline. It supports Claude Code, Gemini CLI, OpenCode, Codex, and Cursor, and is offered free under the MIT license.
Open Source · MIT License · For AI Coding Agents
Engineering Discipline: Guardrails to Stop AI Agents From Spiraling
A free, open-source skill set that structures the workflow from vague requests to verified implementation — embedding planning and independent verification as a defined sequence of steps for tools like Claude Code, Codex, and Cursor.
5
Supported agents:
MIT
Free license,
85★
Stars · 18 forks
THE CORE WORKFLOW
Progressively narrowing a vague request all the way to verified code.
1 · Clarification
Resolve ambiguity, explore the codebase
→
2 · Complexity Assessment
Score routes to Simple or Complex
→
3 · Plan Crafting
Multi-step plan, no placeholders
→
4 · Run Plan
Worker builds, validator verifies
Information Isolation
In execution, a worker implements while an independent validator checks — referencing only the plan and codebase, never the worker's output. Objectivity by design.
WORKERimplements
⊘
VALIDATORverifies blind
Complexity Routing Score
Complex tasks add 5 reviewers + a Long Run Harness.
WHY IT MATTERS
Praised as a fix for the "vibe coding" spiral — declining quality, infinite loops and slop from unstructured prompting. Echoes calls by Karpathy, Osmani, Cherny and Beck for stricter guardrails.
THE CAVEATS
Initial learning cost and context-management limits are noted. No benchmarks or pricing disclosed; the repo updates continuously. Long Run Harness mitigates via checkpoints.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…