Engineering Discipline for
AI Coding Agents

From vague requests → clear plans → verified implementation.
Works directly with Claude Code, Gemini CLI, Cursor, and Codex.

View Workflow ↓ GitHub ↗

01 ▸ Workflow

How it works? Skills are chained to route tasks automatically. All you need to do is say what you want.
User Request (Vague) │ ▼ ┌─────────────────────────────┐ │ clarification │ │ De-risk + Codebase Search │ │ ───────────────────────── │ │ Complexity Check (Routing) │ └──────────────┬──────────────┘ │ ┌────────┴────────┐ │ │ ▼ ▼ Simple (5-8) Complex (9-15) │ │ ▼ ▼ ┌─────────────┐ ┌──────────────────┐ │plan-crafting│ │milestone-planning│ │Plan Creation│ │ 5 Parallel Rev. │ └──────┬──────┘ └────────┬─────────┘ │ │ ▼ ▼ ┌─────────────┐ ┌──────────────────┐ │ run-plan │ │ long-run │ │ Worker- │ │ Orchestration │ │ Validator │ │ │ └──────┬──────┘ └────────┬─────────┘ │ │ ▼ │ ┌─────────────┐ │ │ review-work │◄───────────┘ │ Independent │ Each Milestone: └─────────────┘ plan → run → review │ ▼ [DONE] ─── Independent Skills (On-demand) ─── karpathy → Pre-coding Guardrails clean-ai-slop → Deslop AI output simplify → Quality Review systematic-debug → Root-cause Debugging rob-pike → Measurement-driven Perf
Core Principle: Information Isolation is applied at every transition. Validators never know the worker's intent — only the plan and the result.

02 ▸ Skill Map

Workflow Skills

Clarification
Start Auto-routing

Narrow down vague requests via iterative Q&A + codebase search. Automatically routes to the next skill based on complexity.

Plan Crafting
Plan Executable

Create multi-step implementation plans. Every step must include actual code — no placeholders allowed.

Run Plan
Execute Worker-Validator

Worker implements, Validator independently verifies. Auto-retry on failure, escalation after 3 failures.

Review Work
Verify Isolation

Verify plan vs implementation with zero context overlap. Delivers a binary PASS/FAIL verdict.

Long Running

Milestone Planning
Decompose 5 Parallel Rev.

Synthesize an optimized milestone DAG after concurrent reviews across feasibility, architecture, risk, dependency, and value.

Long Run
Orchestration Checkpoints

Orchestrate executions spanning multiple days. Survives crashes with checkpoints and runs parallel milestones via worktree isolation.

Independent Skills

Karpathy
Prevent Guardrails

Surgical changes, assumption verification, scope discipline — prevent common LLM mistakes before and during coding.

Clean AI Slop
Cleanup 6-Pass

Remove excessive comments, redundant abstractions, and defensive paranoia while preserving behavior. Test-first, single-smell passes.

Simplify
Review 3 Parallel Agents

Reuse, quality, efficiency — 3 parallel reviewers audit changes and automatically fix discovered issues.

Systematic Debugging
Debug 7-Step

Reproduce first → Trace root cause → Failing test → Single fix → Verify. Blocks "fixing other things while I'm here."

Rob Pike's 5 Rules
Perf Measure First

"Don't touch it if you haven't measured it." Prevents premature optimization and enforces data-driven tuning.

03 ▸ Install

Claude Code
/plugin marketplace add tmdgusya/engineering-discipline
/plugin install engineering-discipline
Gemini CLI
gemini extensions install https://github.com/tmdgusya/engineering-discipline
Cursor
/add-plugin engineering-discipline
Codex
npx skills add tmdgusya/engineering-discipline