AI Coding Workflow, Rules & Skills — Complete Guide with Examples

In one line: Many developers treat Cursor prompts as magic spells, but in 2026 the real gap is a three-layer stack — how you Workflow tasks, how Rules lock boundaries, and how Skills package expertise. Below: concepts, architecture, copy-paste examples, and a team rollout checklist. If you already use Cursor or Claude Code, you can scaffold your repo after one read.

Three layers: what Workflow, Rules, and Skills each do

AI coding tools in 2026 (Cursor, Claude Code, GitHub Copilot App, Windsurf) all look like “chat that edits code,” but teams that ship reliably split capabilities into three layers:

• Workflow = tempo on the score (plan first vs. test first)
• Rules = house rules (no backend/ edits, lint before commit)
• Skills = specialists (security review, blog writing, MCP setup)

Layers stack: Workflow shapes the conversation, Rules sit at the bottom of context, Skills inject when the task matches.

Workflow: repeatable collaboration rhythm

Workflow is not a config file — it’s the steps you agree on with the agent.

1. Plan mode (think first)

Best for: cross-file refactors, new features, unfamiliar modules.

Read-only pass: list files to touch, risks, and a step plan with verification per step.
No code until I confirm.

2. Agent mode (small steps)

Best for: approved plan, clear boundaries.

Execute step 1 only under src/auth/. List changed files; do not commit.

3. Review mode (human + agent)

Best for: pre-merge, pre-release.

Review diff vs main: Rule violations, missing tests, hardcoded secrets. Output a checklist.

Rule of thumb: one agent session = one verifiable sub-goal. Don’t bundle “blog + backend + deploy” in one prompt.

Rules: persistent constraints and project memory

Rules mean you don’t repeat “we use TypeScript strict” every session.

.mdc skeleton

---
description: Blog article HTML conventions
globs: frontend/**/blog/articles/**
alwaysApply: false
---

# Blog article rules
- Copy from 0-article-template
- Opening hook before TOC
- Update blog/index.html per locale

• globs: inject only when editing matching files — saves tokens
• alwaysApply: true: global (e.g. “never modify backend”)

Good rules are executable, verifiable, short (200–800 words), and in Git.

Skills: on-demand expertise

Skills are structured capability packs (SKILL.md). vs Rules:

SKILL.md skeleton

# Deploy Skill
## When to use
User asks to deploy, release, or run deploy.sh.
## Steps
1. Confirm no backend/ changes
2. Update seo-file if new blog
3. Run ./deploy.sh
## Forbidden
- No force push to main

Built-in Skills (create-rule, review-security) are good templates.

How layers stack

Priority: current message > Skill steps > Rules > model defaults. Resolve conflicts in text: “Skill covers deploy only; still obey no-backend Rule.”

Example 1: blog Rule

.cursor/rules/blog-article.mdc — globs on frontend/**/blog/articles/**, enforces template copy, hook before TOC, index updates.

Example 2: deploy Skill

.cursor/skills/deploy/SKILL.md — checklist: frontend-only, sitemap/indexnow if blog, ./deploy.sh, verify URL.

Example 3: full blog ship Workflow

Rules: no backend/; Skills inject commands at deploy/sitemap steps.

Team rollout checklist

• [ ] Root AGENTS.md: stack, layout, forbidden areas
• [ ] .cursor/rules/: global + per-directory rules
• [ ] One Skill each for deploy, blog, CR
• [ ] CONTRIBUTING.md: Plan first, PR scope, review checklist
• [ ] Rules/Skills changes via PR
• [ ] Remote machine (Cloud Mac) clones same Git — rules stay in sync

Common pitfalls

1. One giant Rule → context bloat; split + globs.
2. Agent without Plan → wide bad edits; plan first.
3. Duplicating Rules and Skills → single source: constraints in Rules, flows in Skills.
4. “You are an expert” Rules → waste tokens; write checkable rules.
5. Ignoring RAM → multi-agent + builds need 32GB+; offload to Cloud Mac; Workflow unchanged.

FAQ

See structured FAQ in page footer / JSON-LD.

Q: Rules or Skills first?
One global Rule, then 2–3 Skills for top tasks. Workflow emerges from daily use.

Q: Claude Code mapping?
CLAUDE.md ≈ Rules; custom commands ≈ Skills; plan-then-execute ≈ Workflow.

See also Claude Code MCP setup.