Skip to content

Cheat sheet: Developer – Foundations

Everything worth holding in your head the hour before the exam, on one page. Facts come from the official exam guide; the rules of thumb are the maintainer's, distilled from the official rationales.

Claude Certified Developer – Foundations cheat sheet: exam facts, domain weights, and the rules that decide questions

Save or share the card above; the full sheet follows.

The exam
Code CCDV-F
Items 53, multiple choice and multiple response
Time 120 minutes, about 2 minutes per item
Pass 720 of 1,000, scaled
Fee 125 USD before partner discount
Valid 12 months, free renewal assessment

Where the marks are

Applications and Integration        33.1%  ████████████████████████████████
Model Selection and Optimization    16.8%  ████████████████
Agents and Workflows                14.7%  ██████████████
Prompt and Context Engineering      11.0%  ███████████
Tools and MCPs                      10.6%  ██████████
Security and Safety                  8.1%  ████████
Claude Code                          3.1%  ███
Eval, Testing, and Debugging         2.6%  ██

One domain is a third of the paper. Claude Code and Eval together are under 6%, so do not over-prepare them because they feel familiar.

If you remember nothing else

  1. stop_reason drives the loop. tool_use means execute the tool, append the result, call again. end_turn means done.
  2. Batches for latency-tolerant volume. Roughly half the cost, 24-hour window, custom_id correlation, no multi-turn tool calling inside a batch.
  3. Cache the stable prefix. Static system prompt and policy first, variable content last, caching on. Cuts latency and cost together.
  4. Enforce structure, then validate, then retry with the error. Tool use plus JSON schema. Never trust the model to emit valid JSON because you asked firmly.
  5. Known path, workflow. Unknown path, agent. Autonomy earns its complexity only when the model must choose the route.
  6. Subagents isolate context. Delegation keeps verbose intermediate work out of the parent window.
  7. Pin model versions in production. Upgrades then become evaluated changes rather than overnight surprises.
  8. Injection is defeated structurally. Separate untrusted content from instructions and deny it access to consequential tools. Not by asking the model to ignore it, not by a bigger model.
  9. Least privilege on tools. A capability the agent does not hold cannot be misused. Removal beats confirmation dialogs and logging.
  10. MCP server for reuse across applications. Custom tool for one app, Skill for instructions, built-ins for file and shell basics.
  11. Tool descriptions are the selection mechanism. Overlapping, vague descriptions cause wrong tool choices. Differentiate or consolidate.
  12. Context hygiene beats a bigger window. Prune tool output, compact history, isolate subtasks. A larger window delays the same failure at higher cost.
  13. Few-shot fixes format and edge cases. More instructions do not.
  14. Traces localize the fault. Correct reasoning over wrong documents is a retrieval defect, not a model defect.

Decisions the exam keeps testing

Requirement Answer
Cheap bulk work, results tomorrow Message Batches API
Same long preamble every call Reorder for a stable prefix, enable caching
Output must be machine-parseable Tool use with JSON schema plus validation retry
Capability shared by several apps MCP server
Agent must never do X Do not give it the tool
Fixed five-step process Workflow, model used inside it
Long session degrading Prune, compact, delegate

Working the clock

53 items in 120 minutes is a comfortable pace, slightly over two minutes each. My approach:

  • First pass, about 65 minutes. Answer what you know. Code-shaped questions read slowly; do not let one long snippet eat five minutes. Flag and move.
  • Second pass, about 35 minutes. Work the flags. API mechanics questions in particular tend to resolve once you have seen several: the exam is internally consistent about what it considers good practice.
  • Last 20 minutes. Re-read every multiple-response item and confirm the number of selections.

Reading a question the way it is written

Each scenario states a constraint: cost, latency, reliability, security, or maintainability. Identify it in the first sentence, then eliminate. The pattern across almost every item is the same: the correct option changes the structure of the system, and the distractors adjust a parameter, add a bigger model, or ask the model nicely.

Specifically, be suspicious of any option containing "increase max_tokens", "raise temperature", "use a larger model", or "add an instruction to the prompt" when the problem is correctness, cost, format, or security. Those are the exam's standard wrong answers.

The morning of

  • Run the system test again on the exam machine, even if it passed last week
  • Close everything on the shutdown list before check-in, including the Claude desktop app
  • ID ready, room clear, desk clear, second monitor unplugged
  • Check-in takes longer than you expect; start early
  • 135 minutes of seat time; eat first

If you blank

Skip and keep moving. One hard item costs one mark; stalling costs several. On this exam in particular, later items often jog the concept you needed for an earlier one.

Exam day logistics

  • ID name must match your registration exactly
  • Closed book. No notes, no documentation, no translation tools
  • Answer everything; unanswered scores zero
  • Multiple response items say how many to pick
  • Roughly 2 minutes per item; flag and return rather than stalling

Traps

  • The most capable model is rarely the right answer to a cost or latency question
  • Temperature is almost never the fix for a correctness, format, or security problem
  • Try or catch that silently swallows failures is never the sturdy pattern
  • Prompt wording is not a security control, and confirmation dialogs are not least privilege

Study guide · Notes · Practice questions · Mock exam · Repository index