Microsoft AI Dev Days Talk Outline Duration: 15-20 minutes + demos Perspective: The Paradigm Shift Educator
Title Slide: "You've Been Learning the Wrong Thing"
"For 30 years, we've taught developers to learn tools. Memorize the API. Master the syntax. Know the flags. But what if I told you the game just changed—and the skill you need now is something you've been doing since kindergarten?"
Key provocation: The best programmers of the next decade won't be the ones who memorize the most. They'll be the ones who explain the best.
Visual: Side-by-side comparison
- Left:
curl -X POST -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" -d '{"model":"gpt-4"... - Right: "Fetch the weather for Seattle and tell me if I need an umbrella"
Both accomplish the same goal. One requires memorization. One requires clarity.
- Learn the API - Every new tool meant new documentation to read
- Memorize the syntax - Muscle memory for flags, options, patterns
- Master the tool - Deep expertise in specific technologies
- Chain tools manually - Bash pipes, scripts, orchestration
- Cognitive load: Remembering
rsync -avz --delete --exclude='.git'vs. "sync these folders, skip git files" - Context switching: Different mental models for every tool
- Gatekeeping: "You're not a real developer until you know X"
# Show the "old way" - manual tool orchestration
git diff --name-only HEAD~1 | xargs -I {} eslint {} --fix
git add -A && git commit -m "fix: lint errors""This works. But you need to know git, xargs, eslint, and bash piping."
The fundamental shift: We're moving from operating tools to directing tools.
---
command: gh-copilot
---
Look at my recent git changes and fix any linting issues.
Commit the fixes with an appropriate message.- Instructions are the interface - Natural language becomes the API
- Tools become interchangeable - The instruction stays the same; the executor can change
- Composition over memorization - Combine capabilities through description
.md file → frontmatter config → AI executor → result
Why markdown?
- Human-readable
- Version-controllable
- Self-documenting
- Composable
# Create a simple .claude.md file live
ma fix-lint.copilot.mdShow how the instruction file:
- Describes intent clearly
- Works regardless of which files changed
- Could swap
copilotforclaudeorgeminiwithout changing the instruction
Reassure the audience: You're not starting over. Your experience matters.
- Logical thinking - Breaking problems into steps
- Architecture design - Understanding systems and boundaries
- Debugging intuition - Knowing when something's wrong
- Domain knowledge - Understanding the problem space
| Old Skill | New Application |
|---|---|
| Writing clean code | Writing clear instructions |
| Refactoring | Refining prompts |
| Code review | Instruction review |
| Debugging | Validating AI outputs |
| Testing | Verification and guardrails |
Key insight: The mental muscles are the same. The medium changed.
-
Instruction Design
- Clarity over cleverness
- Explicit context setting
- Anticipating edge cases in natural language
-
Escalation Thinking
- When does a task need a smarter agent?
- Cost vs. capability tradeoffs
- Building scripts that know their limits
-
Composition Patterns
- Piping AI outputs to other tools
- Building agent chains
- Non-interactive workflows
---
command: gh-copilot
---
Analyze this error log.
If you can fix it, do so.
If you need more context, output: ESCALATE: <reason>Then show a wrapper script that:
- Runs Copilot first (fast, cheap)
- Detects "ESCALATE" in output
- Pipes to Claude (slower, smarter) automatically
This is the paradigm shift in action: Software that knows when it needs help.
-
Write instructions, not just code
- Practice explaining tasks to AI like explaining to a junior dev
- Build a library of instruction templates
-
Think in capabilities, not commands
- "I need to transform data" not "I need to learn jq"
- The tool is discovered, not memorized
-
Design for composition
- Non-interactive modes enable piping
- Structured outputs enable chaining
- Exit codes enable conditional flows
-
Build verification habits
- AI outputs need validation
- Trust but verify becomes the mantra
- Automated checks > manual review
# Show piping in action
ma analyze-code.copilot.md | ma create-tests.claude.md | ma review.copilot.mdThree agents, each with a clear job, composed through pipes.
"What if the next tool you need to learn... is how to explain what you want?"
- Write one instruction file - Take a repetitive task and express it in markdown
- Practice escalation - Build a script that tries the cheap option first
- Pipe an AI - Use non-interactive mode and chain the output somewhere
Old: "I need to learn this tool's API" New: "I need to clearly express what I want"
Old: "What command does this?" New: "What instruction accomplishes this?"
Old: "Master the tool" New: "Master the explanation"
"The developers who thrive in this new world won't be the ones with the best memory. They'll be the ones with the clearest thinking. And that's actually great news—because clear thinking is a skill anyone can develop."
| Demo | Purpose | Time |
|---|---|---|
| Traditional bash pipeline | Show the "old way" complexity | 1 min |
| Simple .md instruction file | Introduce the new paradigm | 2 min |
| Intelligent escalation script | Show agents knowing their limits | 3 min |
| Live agent composition | Demonstrate piping AI outputs | 2 min |
Total demo time: ~8 minutes Total talk time: ~12 minutes Buffer for transitions/Q&A: ~5 minutes
- Software is now instructions + tools - The interface is natural language
- Your experience transfers - Logic, architecture, debugging still matter
- New skills to cultivate - Instruction design, escalation thinking, composition
- The shift is from memorization to explanation - Clarity is the new expertise
- Start now - Write instructions, practice escalation, pipe your AI