Common First-Timer Mistakes
Pitfalls to avoid when starting with ASD
New practitioners consistently fall into the same traps. Recognizing these patterns early prevents frustration and builds effective habits faster.
Skipping the planning phase
Agents perform poorly when given vague, open-ended requests. "Make this better" or "add a login system" without context forces the agent to guess at requirements.
The solution is explicit planning: ask the agent to outline its approach before generating code, then review and adjust the plan. Claude Code's Plan Mode (accessed via Shift+Tab twice) exists specifically for this purpose.
Treating agent output as production-ready
Studies consistently find that 60-70% of agent-generated code requires modification. First-timers often accept output without critical review, leading to subtle bugs and security vulnerabilities.
Every generated code block requires the same scrutiny applied to a junior developer's pull request.
Overloading context
The 200k token context window does not mean stuffing 200k tokens improves results. Research shows agents perform worse with excessive context due to "lost in the middle" effects where information between the beginning and end receives less attention.
Effective practice involves providing targeted, relevant context rather than dumping entire codebases.
Misunderstanding the developer role
ASD requires shifting from writing code to directing and validating. First-timers often fight this transition, attempting to code alongside the agent rather than orchestrating its work.
The architect, director, and reviewer roles discussed earlier represent genuine skill sets requiring deliberate practice.
Ignoring tool differences
Claude Code and Codex have fundamentally different interaction models:
| Claude Code | Codex |
|---|---|
| Conversational flow | Transactional approvals |
| Extended dialogue | Discrete actions |
| Collaborative iteration | Sandbox-first security |
Understanding each tool's strengths guides appropriate task assignment.