Method · 3 steps
The 3-step method
How I read every question from the last week of prep onward, and two rules for when two options still look right.
The idea behind most questions
Most questions turn on one distinction: advisory versus enforced. Anything that lives in the model's context is advisory: a system prompt line, a CLAUDE.md note, a few-shot example. The model will usually follow it, and "usually" is the failure the question describes. Enforced means something outside the model's choice makes it happen: a forced tool_choice, a PreToolUse hook that denies a call, a check in the tool handler, a schema, a path-scoped rule that loads on its own. When a question says guarantee, must, never or always, look for the enforced option.
Three steps
Step 1 · Name the failure in one line
"The loop stops while a tool call is pending." "The rule loads in sessions that never touch the files." If you can't write the line, reread the scenario.
Step 2 · Find where the cause lives
It is almost always one of five places: the prompt, a tool (its description, its errors), code (the loop, a handler, a validator), configuration (CLAUDE.md, rules, hooks, skills, flags), or a handoff between agents.
Step 3 · Pick the option that changes that place
Not the one that adds a layer somewhere else.
Two tie-break rules
Rule 1
Keep your first answer
Keep it unless you can point to a specific fact in the question that it ignores. Switching to the answer that feels more careful cost me more points than anything else.
Rule 2
Prefer the smallest enforced change
Prefer the option that acts at the source, is enforced rather than advised, adds nothing new (no extra agent, reviewer or service), and uses only names you have seen in the docs. A flag or setting you have never seen is usually invented.
Traps that cut across domains
- The heavier-sounding option. An extra agent, an extra review layer, a human flag, a longer prompt. In my practice runs this was the most common miss. The answer was the smallest change at the cause.
- A familiar word on a different layer.
context: forkis not a git fork. Forking a session is not a git worktree. A "hook" in copywriting is not a Claude Code hook. When a word feels familiar, check which layer it belongs to in this question. - Reading state from the shape of a response. "There's a text block, so it's done." "It repeated the conventions back, so they loaded." Use the field or command that reports the state directly.
- Trusting the aggregate. 96% overall can hide one document type at 70%. Look for the option that breaks the number down.
- Dropping a correct first answer because of an unfamiliar term. Work out what the term means before you let it move you.
How I studied
- Diagnostic first. Take a timed practice set before you study, so you know your weak domains. Mine were prompt engineering and Claude Code configuration.
- Domain by domain. For each domain, read the task statements in the exam guide and the matching Anthropic docs. Put every term you don't know into a glossary the moment you meet it. Test the domain with a couple of questions before moving on.
- Drills, logged. Short sets, and one line per miss: what was missing. After a while the misses group into a few patterns, and the patterns are what you study. If the right answer in your drill bank keeps being the same letter or the longest option, the bank is leaking answers.
- A timed mock in the middle, not at the end. Leave time to act on what it shows. Do a second one as a rehearsal.
The habit that kept it going was a two or three minute warm-up at the start of each session: reread the traps and the advisory-versus-enforced idea, nothing else. Once a week, a longer pass through the logged misses and a few questions cold.