Skip to content

claude_quick_reference.md

Latest commit

 

History

History
112 lines (79 loc) · 4.17 KB

claude_quick_reference.md

File metadata and controls

112 lines (79 loc) · 4.17 KB
Raw

Getting Results with Agentic AI: Quick Reference

The Key Difference from Copilot

Copilot Agentic AI (Claude Code, Cursor, Windsurf)
You write code, AI suggests AI writes code, you steer
Tab to accept Conversation to collaborate

Key shift: You're directing an agent, not accepting suggestions.

What Makes Sessions Effective

Principle How
Be Detailed "I see X when I expected Y" not "it's broken"
Validate Assumptions After work: "Did you write tests first?" "What did you understand?"
Close the Loop Report results: "yup", "I see X", "that worked"
Steer Constantly Direct frequently: "Great. Now let's...", "OK, let's commit"
Redirect Gently Soft corrections: "oh wait", "actually" (not "no", "wrong")

The Power of Detail

Detail helps at every stage:

Context Instead of... Say...
Problems "It's broken" "I see white screen after JS bundle loads"
Feedback "It works" "yes! but the cancel button is getting occluded"
Corrections "Wrong file" "oh wait, I meant the production config"
Steering "Add error handling" "let's add an error boundary so users can share errors"

The 8 Patterns

Pattern Signal Words Example
Feedback Loop "I see", "yup", "that works" "yup that worked"
Quick Question short + "?" "should we commit first?"
Thinking Aloud "...", "hm,", "well.." "well.. do we need this?"
Affirm + Pivot "Great.", "nice." + next "nice. let's write tests"
Course Correct "oh", "actually", "wait" "oh wait, not that one"
Context Dump Problem + tried + seeing [detailed opener]
Layer Context "btw", "fyi", "also" "btw, this needs to work offline"
Directive + Why action + "because" "let's add docs for future devs"

Session Arc

SETUP (10%) → BUILD (20%) → EXECUTE (30%) → ITERATE (25%) → FINISH (15%)
   Context      Questions     Deep work       Debug         Cleanup
   Alignment    Iteration     Steering        Refine        Verify

Key: Direction messages ("let's do X") appear 30-40% in ALL phases.

You're not coding—you're navigating. Stay in the driver's seat.

Know Your Starting Point

If you... Then...
Have a fuzzy idea Explore together: "I'm thinking about X, what are the tradeoffs?"
Have something complex Start in plan mode: /plan - get alignment before code
Know what you want Execute efficiently: "I wrote it up in file.md, let's build"
Get stuck mid-session Step back: "Let me clarify what I'm actually trying to do here"

Plan mode: Claude Code has /plan, Windsurf has Q&A mode. For other tools, prompt "ask clarifying questions before starting." Worth the upfront investment for anything non-trivial.

What to Avoid

Problem Fix
Frustrated circling ("try again", "still not working") Stop and diagnose—do you actually know what you want?
AI keeps misunderstanding Write context to a file, point to it
Bare "yes" or "ok" Add context: "yes, and also..."
Hard corrections ("no", "wrong") Soft redirect: "oh wait", "actually"
Staying silent after actions Report results (even "yup")
Letting AI run unsupervised Steer every few exchanges

Getting Started

  1. Start small - Single file, not whole feature
  2. Close the loop - Always report what happened
  3. Think out loud - Share your uncertainty
  4. Steer constantly - "Let's do X next"
  5. Write context to files - For anything complex

The Bottom Line

Effective sessions aren't about fewer turns—they're about making progress toward your goal.

  • Exploring a fuzzy idea? More back-and-forth is fine. That's the point.
  • Building something clear? Write it down first, then execute efficiently.
  • Getting frustrated? Stop and diagnose. Something's wrong with the setup.

Stay in the driver's seat. Steer constantly. Report results. That's how you get results.