Skip to content

AI Workflows

Practical guide: what to use, why, tradeoffs, and a minimal setup.

Quick Decision

  • One-off task -> Prompt
  • Repeatable behavior rule -> Instruction
  • Repeatable task flow -> Skill
  • Multi-step execution -> Agent
  • External system/tool access -> MCP
  • Fast shell/devops flow -> CLI

Topics

Prompt

  • Best for: One request
  • Benefit: Fast and flexible
  • Drawback: Quality depends on context quality

Instructions

  • Best for: Team conventions
  • Benefit: Consistent outputs across sessions
  • Drawback: Can become noisy if too broad

Skills

  • Best for: Repeatable workflows
  • Benefit: Reusable, structured execution
  • Drawback: Upfront design effort

globally installed skills are located here: %USERPROFILE%\.agents\skills or %USERPROFILE%\.copilot\skills

Agents

  • Best for: Complex multi-step tasks
  • Benefit: Better planning and autonomy
  • Drawback: More setup and maintenance

MCP

  • Best for: External tools and APIs
  • Benefit: Real data and integrations
  • Drawback: Security and ops overhead

Loading Behavior

  • Instructions are not automatically loaded in every case.
  • Broad instructions can be loaded very often, for example global instructions or wide applyTo patterns.
  • Narrow instruction files are only useful if they match the current task.
  • Skills are usually loaded on demand, not on every request.
  • Result: put stable behavior rules in instructions, but put heavier workflows into skills or agents.

Minimal Setup

  1. Add global behavior rules in AGENTS.md or .github/copilot-instructions.md.
  2. Add domain instructions for focused areas (api, security, docs).
  3. Add one prompt for recurring asks (for example: "draft ADR", "review API").
  4. Add one skill for a common workflow (for example: "create concise doc page").
  5. Add one agent when task orchestration becomes repetitive.

Example structure:

.github/
  copilot-instructions.md
  instructions/
    docs.instructions.md
    api.instructions.md
  prompts/
    write-short-doc.prompt.md
  skills/
    create-doc-page/SKILL.md
  agents/
    docs-editor.agent.md
AGENTS.md

Concrete Usage

  • Prompt: "Summarize this page in 5 bullets with 1 action per bullet."
  • Instruction: "For docs pages: short headings, short bullets, avoid duplicate content."
  • Skill: "When asked to create docs, first link related pages, then draft minimal page."
  • Agent: "Investigate, propose options, apply edits, run checks, report risks."

Token Optimization

  • Keep instructions short. Long global instructions cost tokens on many requests.
  • Avoid applyTo: "**" unless the rule really applies everywhere.
  • Split rules by area: docs, backend, frontend, security.
  • Link to existing docs instead of embedding the same guidance again.
  • Prefer small prompts with good context over large generic prompts.
  • Use skills for larger workflows so the details are loaded only when needed.
  • Remove stale instructions. Old rules create noise and contradictions.

Good for tokens:

  • "Write a short page in the style of docs/frontend/best-practices.md."

Bad for tokens:

  • Large prompt + copied style guide + copied examples + repeated repo rules

Typical Workflow

  1. Start with a prompt and success criteria.
  2. Promote repeated guidance into instructions.
  3. Promote repeated process into a skill.
  4. Promote repeated orchestration into an agent.
  5. Add MCP only if local context is not enough.

Best Practices

  • Keep each instruction file narrow in scope.
  • Prefer links to existing docs over copied text.
  • Treat skills as small building blocks.
  • Keep agents focused on one workflow outcome.
  • Review AI output before merge or publish.

Learnings

  • Use simple models for simple tasks, use heave models (expensive one) for complex tasks. Judging simple and complex is always a gut feeling, one question to help however could be "How many decisions need to be made?" -> many decisions -> complex
  • Use *.instructions.md files for general AI guiding or team wide configurations
  • Use SKILLS.md for specific actions to execute. Skills can be auto discovered based on there meta data (description) and will only be loaded into context when it matches the current task at hand. They might load additional content (such as references (templates)) in cases their execution need it.