Working Effectively with AI Agents: What I’ve Learned Using Claude
I wanted to take a moment to share something small, but valuable for anyone working deeply with AI agents.
After extensive use of Claude, one thing has become very clear:
You cannot treat AI like a tool you trust blindly. You have to treat it like a capable assistant that requires strong supervision.
That shift in mindset changes everything.
1. Don’t Trust the Output—Verify It
Never assume the work is correct.
You need verification built into your workflow:
- Independent checks
- Testing (E2E, UAT, etc.)
- Adversarial review
If you don’t actively verify, you will eventually ship something incorrect.
2. Your Role Is the Supervisor
Claude performs best when you act as a clear, disciplined operator.
That means:
- Reviewing everything it returns
- Catching when it drifts from instructions
- Reinforcing expectations consistently
It will make mistakes. That’s part of the system. Your job is to guide it back on track.
3. Control Context—More Is Not Better
One of the biggest mistakes is overloading context.
Give Claude exactly what it needs for the current step—no more.
Too much context introduces noise, confusion, and lower-quality outputs.
4. Give It Tools—and a Way to Check Itself
Claude is significantly more effective when it can:
- Research
- Execute
- Validate its own work
Whether that’s automated testing, CLI tools, or structured verification loops—you want it operating in a system where it can double-check itself.
5. Always Work from a Task System
If you don’t have a clear, sequential task list, things fall apart.
A persistent task system ensures:
- Work is completed in order
- Nothing is skipped
- You always know what’s next
Without it, you will lose direction.
6. Establish a Working Contract
Define how the agent operates.
For me, that includes:
- No claiming success unless the work is actually complete
- No hiding mistakes—own them
- Capture “gotchas” and lessons learned
- Prioritize accuracy over speed
This belongs in your operating file (e.g., claude.md), and it should be explicit.
7. Plan First, Build Second
Take the time to think.
In practice, this means:
- Spending significantly more time planning than coding
- Breaking work into phases when needed
- Asking for an MVP when appropriate
Interestingly, I’ve found it’s often easier to have Claude build something new than fix something poorly thought out.
8. Communication Matters More Than You Think
Garbage in, garbage out.
Be:
- Clear
- Direct
- Structured
And just as important—stay constructive. Frustration doesn’t improve results.
9. Stay Aligned on the Big Picture
You bring the vision.
Claude brings the execution capability.
When both are aligned, progress is fast. When they’re not, you get drift.
10. Use It to Find What You Don’t See
One of the most valuable uses of AI is identifying blind spots.
But here’s the catch:
It won’t surface them unless you ask.
Prompt it to:
- Challenge your assumptions
- Identify risks
- Highlight what might be missing
That’s where real leverage comes from.
11. Small Inputs Can Unlock Big Ideas
Sometimes a single concept or keyword, combined with the right prompt, can lead to something entirely new.
AI is extremely powerful at synthesis—but only if you feed it the right signals.
12. Handle Continuity Intentionally
Each session is effectively a new instance.
That means:
- Maintain handoff summaries
- Keep memory lightweight
- Store detailed context externally
Your system—not the model—owns continuity.
13. Keep Core Files Lean
Your main instruction files should be minimal.
- Under ~100–200 lines
- Focused on principles and structure
- Pointing to external resources when needed
Overloading them wastes tokens and reduces effectiveness.
14. Separate Sensitive and Operational Data
Don’t mix everything into a single file.
Keep structured separation:
- Global instructions
- Task tracking
- Private or sensitive information
This keeps your system clean and scalable.
Final Thought
Claude is not just a tool.
It’s an extension of how you think and operate.
If you’re structured, disciplined, and clear—you’ll get exceptional results.
If you’re not, the system will reflect that just as quickly.
Smooth is fast.