9.1 — Agent vs chatbot: a posture shift
When you use ChatGPT or Claude in a web interface, you're interacting with a chatbot. You have to feed it all the context -- the code, the files, the architecture -- because it can't fetch information itself. Claude Code is an agent. It has tools: Read, Grep, Glob, Bash, Edit. It explores your codebase, reads your files, runs commands, navigates the git history. This difference radically changes how you should phrase your instructions.
With a chatbot, you give the steps. With an agent, you give the direction. Instead of saying "open file X, find line Y, replace Z with W", you say "date sorting is broken on the orders page, fix it". The agent will find the file, understand the bug, propose the fix. Your energy should be invested in the "what" and the "why", never in the "how".
Concretely:
- Don't paste code into your prompt -- the agent can read it itself
- Describe the symptom and the expected behavior, not the technical steps
- Think "product owner writing a ticket", not "developer dictating code"
- Trust the exploration: the agent will find the files, patterns, and architecture
# Bad (chatbot style)
"In src/utils/sort.ts, line 42, replace a.date - b.date with b.date - a.date"
# Good (agent style)
"Date sorting in the orders list shows the oldest first.
I want the most recent first."
Key takeaways
Prompting an agent means delegating to a competent developer. You give the problem, the business context, and the constraints. You don't dictate every line of code.