On this page · 7 sections
In the summer of 2026 I studied for the Claude Certified Architect – Foundations exam. I passed. I’d used Claude Code daily for months, but I’d never called a model through an API or used an SDK. I gave it two hours a day. This isn’t a post about exam questions. It’s about the ideas that stuck, because they changed how I build.
TL;DR
- When a failure is deterministic, change the tool or the schema, not the prompt.
- A PreToolUse hook can block a tool call; a PostToolUse hook can only react.
- A coordinator splits work and combines results. Subagents get only the context they need.
- Validate single fields and cross-field rules separately, and return structured errors.
- Few-shot examples fix variation in phrasing, not a rule the model keeps breaking.
Fix the structure, not the prompt
The biggest lesson. When a failure is deterministic (the agent always picks the wrong tool, always misses a field), a better prompt is the weakest fix. Change the interface instead: a clearer tool description, a required parameter, a schema that rejects bad input. Save the prompt for nuance.
If it must always happen, don’t ask for it. Make it the only thing that can happen.
Know when hooks run
Claude Code’s hooks fire at specific moments. A PreToolUse hook runs before a tool call and can block it (one of mine does). A PostToolUse hook runs after and can react to the result. I mixed these up early. Putting a check in the wrong one means it either can’t stop anything or runs too late to matter.
Coordinators coordinate
In multi-agent designs, the coordinator splits the work, passes context and combines results. Subagents do focused jobs with only the context they need. When a coordinator starts doing the work itself, or a subagent needs the whole conversation, the design is wrong.
Validation has two levels
Some checks belong to one field (is this a valid date?). Others need several fields together (is the end date after the start date?). Put each where it belongs, and return structured errors the agent can act on rather than prose it has to interpret.
Few-shot examples are for patterns
Examples work best when the problem is variation in surface form: the same idea phrased many ways. They’re a poor fix for a rule the model keeps breaking. That goes back to structure.
How I’d study again
- Build small things against the API from week one, not just read.
- Do practice questions early to find gaps, then study the gaps.
- Argue with the answer keys. Some mock answers were wrong, and working out why taught me more than the right ones.
- Write down every mistake as a one-line rule. That list became my revision notes, and later, half of this post.

