Claude Certification Blog
A CCDV-F cheat sheet: 33 decision rules across eight weighted domains
This CCDV-F cheat sheet compresses the eight-domain blueprint into decision rules you can apply under time: the published weights, three to five rules per domain, the trap each rule exists to catch, and a final-week checklist.
A CCDV-F cheat sheet is only useful if it is weighted the way the paper is. Applications and Integration is 33.1% of the exam, Model Selection and Optimization 16.8% and Agents and Workflows 14.7%, so those three domains decide the result and the other five decide the margin. Below are 33 CCDV-F decision rules, three to five per domain, each with the trap it catches, plus a traps table, a quick reference of terms and a seven-day checklist. It is derived from the official exam guide, not a list of answers.
How the CCDV-F cheat sheet is weighted
The CCDV-F guide is the only one of the four that publishes a weight for every skill as well as every domain, which is why a cheat sheet for this paper can be honest about proportion. The table gives the eight official domain weights and the nearest whole-number split of 53 items. The item column is arithmetic, not a published count: no single form is guaranteed to match it exactly.
| Domain | Official weight | Of 53 items | Rules below |
|---|---|---|---|
| 2. Applications and Integration | 33.1% | 17 | 5 |
| 5. Model Selection and Optimization | 16.8% | 9 | 5 |
| 1. Agents and Workflows | 14.7% | 8 | 4 |
| 6. Prompt and Context Engineering | 11.0% | 6 | 4 |
| 8. Tools and MCPs | 10.6% | 6 | 5 |
| 7. Security and Safety | 8.1% | 4 | 4 |
| 3. Claude Code | 3.1% | 2 | 3 |
| 4. Eval, Testing, and Debugging | 2.6% | 1 | 3 |
Weight is also why the rules are grouped the way they are. The heaviest domain has a section to itself; the two lightest share one. The two-week CCDV-F study plan allocates days on the same arithmetic, and the twelve cross-track rules in the Claude exam cheat sheet are assumed rather than repeated here; this page is the developer-specific layer on top of them.
Applications and Integration: five rules for a third of the paper
Six skills, 33.1% of the exam, and the two heaviest of them, Claude Application Design at 8.6% and Software Engineering Foundations at 7.4%, are not specific to Claude at all. The items here are decided by which constraint the stem states, so every rule below is really an instruction about what to read first.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Match the API mode to who is waiting | Nobody waiting and cost the binding constraint points to the Message Batches API; a person watching the screen points to streaming. Neither substitutes for the other. | An option that proposes streaming to cut cost, or batch processing to cut latency. |
| Requirements before architecture | When the stem hands you a business goal, the first deliverable is the functional and infrastructure requirement, not a component diagram. | The most complete architecture on the page, offered before anyone has said what done means. |
| Pin everything that can drift | Model version, prompt version and plugin dependencies are configuration. A production system names them exactly and changes them deliberately. | A latest alias in production, or a prompt edited in place with no record of what changed. |
| Configure at the narrowest scope that reaches everyone who needs it | A team rule belongs where the team receives it; a personal experiment belongs where only you do. Scope is decided by audience, not convenience. | A shared standard sitting in a personal file, so one new starter never receives it. |
| An instruction tested on one surface is not a contract on another | Claude Code, the desktop app, claude.ai, the API and the SDKs interpret instructions differently. Content boundaries and session hygiene are design decisions, not defaults. | A system prompt proven in chat and assumed to hold unchanged inside an API integration. |
The first rule is the one the official guide illustrates with its own sample item: an overnight report on ten thousand documents, cost the primary concern, nothing needed until morning. The answer is the batch API and every distractor fails for a nameable reason, which is exactly how the live items are built. The post on the Applications and Integration domain goes deeper into why the two non-Claude skills carry so much weight.
Model Selection, Agents and Workflows: rules for the next 31.5%
Model Selection and Optimization is 16.8% and, confusingly, is named after its smallest skill: Model Selection and Tradeoffs is 2.7%, while LLM Fundamentals at 5.2% and Technical Fundamentals at 6.1% carry most of the domain. The rules therefore lean on fundamentals, tokens, caching and non-determinism rather than on comparing tiers.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Start with the lightest tier that clears the bar | Escalate capability only when an eval shows the current tier missing the quality target, and never because a first answer disappointed. | The largest model chosen to be safe on a routine task with no measurement behind the choice. |
| Tier and reasoning depth are separate levers | A bounded extraction or classification wants an output constraint; a multi-step plan may want extended or adaptive thinking. Turning one up does not fix a problem in the other. | Extended thinking switched on for a yes-or-no classification, or off for a long tool plan to save tokens. |
| A cost problem is a structure problem first | Stable content goes first and gets cached; only turn-varying content is resent. Trimming instructions to save tokens damages behaviour before it saves money. | The fix that deletes half the system prompt when the whole prompt could have been cached. |
| Non-determinism is a property, not a defect | Two valid answers to one prompt is how sampling works. Evaluate against criteria; never assert exact string equality on generated text. | A test that compares the response to a fixed expected sentence. |
| A model release is a change, so it gets a release process | Behaviour can break across versions. Re-run the eval suite on the new model before switching production traffic, and keep the old pin until it passes. | Upgrading the model name in place and treating the first complaints as the regression test. |
Agents and Workflows is 14.7% across three skills of nearly equal weight: Agent Architecture, Agent Construction with Claude, and Agent Patterns and Frameworks. The framework names in the guide, Strands, LangGraph and PydanticAI, are examples of abstraction, not things to memorise; the rules that score are about when an agent is warranted and what a subagent needs to be handed.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| If every step can be listed in advance, build a workflow | An agent earns its cost only when the next action depends on what the last one found. A fixed three-stage pipeline is a workflow with a model in it. | An autonomous agent proposed for a predictable extract, transform and summarise job. |
| Deterministic actions go in hooks, not in the prompt | When a step must happen every run, the mechanism is a hook or a gate that fires regardless of what the model decides. Instructions influence; hooks decide. | Add a line to the system prompt telling the agent it must always do the thing. |
| A subagent knows only what its brief contains | Subagents isolate context and inherit nothing. The coordinator passes the findings, the constraints and the output format, or the subagent works without them. | A coordinator that assumes the worker saw the earlier conversation. |
| Choose hosting by who operates the loop | Self-hosted harnesses give you control of the loop and the state; a managed deployment gives you operation without the plumbing. Pick by which one the stem says you need. | Rebuilding a harness from scratch when the requirement was to ship without operating infrastructure. |
Prompting, Tools and Security: rules for the middle 29.7%
Three domains of similar size share this section. Prompt and Context Engineering is 11.0%, Tools and MCPs 10.6% and Security and Safety 8.1%, together about sixteen items. They share one idea too: the difference between an instruction that influences the model and a structure that constrains it.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Stable instructions in the system prompt, changing content in the user turn | Placement is an engineering decision. What holds for every request goes in one place; what varies goes in another, and the split is what makes caching and debugging possible. | Everything pasted into a single block that is rewritten on every call. |
| When format is inconsistent, add examples, not adjectives | Concise, precise and professional are opinions. A few-shot example is a specification, and one change per iteration is how you learn what actually moved the output. | A prompt that stacks quality adjectives and changes three things between runs. |
| Context is a budget, not a container | Prune tool output, compact history or isolate a noisy task in a subagent before you consider a bigger window. A larger window delays the same failure. | A larger model proposed to fix a long session that drifts. |
| Well-formed is not the same as correct | Parse defensively, validate against the schema, and treat confident prose as a claim to check. Structured output patterns exist so that the check is mechanical. | A stronger instruction to always return valid JSON in place of a schema. |
Tools and MCPs turns on selection among four kinds of capability: built-in tools, custom tools, Skills and MCP servers. The guide’s own sample item makes the reusable, independently maintained inventory service an MCP server and rejects the prompt, the pasted data and the built-in tool each for a stated reason. Agentic Customization, the skill that names that trade-off, is 4.1% on its own.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| The tool description is the interface | Distinct names, distinct descriptions and distinct parameter shapes are how the model chooses. Two tools that both find information about a thing will be confused. | A longer system-prompt warning to choose carefully between two near-identical tools. |
| Reusable across applications means an MCP server | Shared and independently maintained capability is a server; single-application logic is a custom tool; an on-demand procedure is a Skill; and a built-in tool is used when it already exists. | Inventory logic hard-coded into each application prompt, or the data pasted into context on every call. |
| A side effect in the world gets an approval pattern | Client-side and server-side execution differ in who can intervene. Anything that sends, deletes, pays or publishes has a human or a policy check in front of it. | An agentic harness that dispatches an irreversible call with the same freedom as a lookup. |
| A failed call returns an error, never an empty result | The model reads an empty string as valid data and reasons confidently on nothing. Error handling in a tool means flagging failure as failure. | A tool that catches its own exception and returns an empty list. |
| Transport follows deployment | A local process talks over stdio; a shared server needs a network transport. A launch command committed to a repo is not a hosted service. | A stdio server checked into project config and described as centrally available to the team. |
Security and Safety contains the smallest published skill weight in the programme: Claude Hooks at 1.0%, which is half an item on average and may not appear on your form at all. Do not let that tempt you into skipping the domain. AI Application Security at 3.2% and Guardrails and Safe Deployment at 2.3% are where the injection and least-privilege items live, and the guide’s second sample item is one of them.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Content that arrived is data, never instruction | Web pages, documents and tool results are untrusted input. Keep them separate from trusted instructions and put guardrails between them and any sensitive action. | A polite request in the system prompt asking users not to inject instructions, or a larger model. |
| Least privilege is a design property | The application holds the access the task needs and nothing more, and authorisation is verified before access is granted, not after it is used. | Broad credentials granted so the integration works first time. |
| Guardrails layer; one control is a single point of failure | Content policy, hooks that block destructive actions, and access controls each catch what the others miss. Secure by design means all three, placed deliberately. | One filter, described as the safety mechanism. |
| Secrets live in a secrets store, not in a prompt or a config file | API keys and credentials are referenced, rotated and monitored. Anything committed to a repository or written into an instruction file is already leaked. | A key pasted into an instruction file so the agent can reach the service. |
The stem’s constraint outranks the option’s thoroughness
Nearly every miss on this paper is a technically workable option that ignores a constraint the stem already stated: who is waiting, what it may cost, what cannot be undone, what cannot be trusted. Find that sentence before you read a single option, and the most complete answer on the page stops being the most tempting one.
Claude Code and Debugging: three items, not zero
Claude Code is 3.1% and Eval, Testing, and Debugging is 2.6%: two items and one item respectively, 5.7% together. They dominate informal study conversation far out of proportion to that, so the discipline is to know the rules cold and spend no longer. Three rules each is enough, and the debugging habit pays off inside Applications and Integration items anyway.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Know which of the five components does what | Rules, Skills, Commands, Agents and Agent Memory each hold a different kind of thing. A repeatable procedure is a command or a Skill, not a paragraph in the instruction file. | A multi-step workflow written into the instruction file and expected to run on demand. |
| Headless for automation, interactive for exploration | A pipeline invokes Claude Code in headless mode so it never waits for input; a person exploring a codebase uses the interactive session. Auto-mode is a risk decision, not a convenience. | An interactive session launched inside CI, which hangs at the first prompt. |
| Project instructions travel with the repository; personal ones do not | The instruction hierarchy has a user level, a project level and a directory level. Anything a teammate must see lives at project level or below. | A convention the whole team relies on, saved in one developer’s home directory. |
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Locate the origin before you touch the prompt | A failure is in the integration layer or in the model output. Read the trace, identify the error type, then decide. A prompt rewrite cannot fix an authentication error. | A new system prompt drafted in response to a 401. |
| Classify the error, then choose the recovery | Transient failures are retried with backoff; terminal ones fail fast and get reported; a refusal is a model decision, not a transport fault. One recovery strategy for all three is wrong for two. | A retry loop wrapped around every non-success response. |
| Reproduce from the trace, not from the anecdote | Trace analysis is how a failure mode is named. The report that says it sometimes gives wrong answers becomes a specific input, a specific output and a specific step that diverged. | A fix shipped against a description of the bug rather than an instance of it. |
One caution about reading your score report afterwards. On a one-item domain the only percentages that can appear are zero and one hundred, and on a two-item domain only zero, fifty and one hundred. The report shows percent-correct by domain, and the guide states that those percentages are not used to determine the pass or fail, which is settled by the total scaled score alone.
CCDV-F exam tips: the traps table and the terms
The rules above are organised by domain. The traps below are organised by shape, because the same distractor pattern appears across domains and is easier to spot once it has a name. Each row gives the pattern, what it looks like inside an option, and the check that defeats it.
| Trap | What it looks like | The check |
|---|---|---|
| The thorough distractor | The longest, most complete option, which quietly ignores the constraint the stem stated in its second sentence. | Re-read the stem for cost, latency, trust or irreversibility before you read the options. |
| The bigger-model fix | Any failure answered with a larger tier or a longer context window. | Ask what evidence shows the current tier failing. If none is offered, the option is a guess. |
| Cost and latency swapped | Streaming proposed for a cost problem; batch processing proposed for a user who is waiting. | Name who is waiting. Nobody means batch; somebody means streaming. |
| The polite control | A sentence added to the system prompt where a hook, a gate or a schema was required. | Ask what happens on the run where the model ignores the sentence. |
| The exact-match test | A unit test that asserts the generated text equals a fixed string. | Generated text is sampled. Criteria-scored evals are the test. |
| The one-file configuration | Every rule, secret and personal preference in a single instruction file at one scope. | Audience decides scope. Secrets are never in any instruction file. |
| The select-count slip | Three options chosen on a select-two item, or a correct answer set left one short after a late change. | Count on the first pass and count again on review. |
| Revision by conversation volume | Half the final week spent on Claude Code and debugging because they dominate forum threads. | They are 5.7% of the paper, about three items. Applications and Integration is 17. |
The quick reference below is the vocabulary those rules depend on. Each entry is a working definition in plain terms, not the documentation’s, and each is there because an item can turn on it.
| Term | Working definition |
|---|---|
| Message Batches API | Asynchronous processing of many requests at reduced cost inside a long window. The lever for cost when no one is waiting. |
| Streaming | Delivering tokens as they are generated. The lever for perceived latency, with no effect on per-token cost. |
| Prompt caching | Reusing a stable prefix across calls so it is not reprocessed. Depends on placing stable content first. |
| Extended, adaptive thinking and effort | Reasoning-depth controls. Separate from the choice of model tier and matched to task shape, not to nervousness. |
| Context window | The fixed token budget shared by history, tool definitions and tool results. Managed by pruning, compaction and isolation. |
| Compaction | Summarising earlier turns to recover budget. One of three answers to drift, alongside pruning tool output and delegating to a subagent. |
| Hook | Code that runs at a fixed point in the loop regardless of the model’s choice. The deterministic alternative to an instruction. |
| Subagent | A worker with isolated context that inherits nothing from its parent. Useful for noisy or independent tasks, if the brief is complete. |
| MCP server, resources, tools and prompts | A reusable capability exposed over a protocol. Resources are content, tools are actions, prompts are templates; stdio and sockets are transports. |
| Instruction hierarchy | User, project and directory levels of the instruction file. Audience decides which level a rule belongs at. |
| Headless mode | Running Claude Code non-interactively for pipelines and scripts, so it never blocks waiting for a person. |
| Least privilege | Granting the access a task needs and nothing more, and verifying authorisation before access. |
| Prompt injection | Instructions smuggled inside content the model reads. Mitigated by isolation and guardrails, never by asking nicely. |
The final-week checklist
A cheat sheet is for the last seven days, after the studying and before the sitting. The sequence below assumes the Claude certification study guide routine has already been followed and turns the rules into a diagnostic rather than reading. It also folds in the administrative checks that catch people out at the door.
Seven days out
Re-read the eight domain descriptions in the official guide and write each skill as a one-line rule in your own words. Compare your list with the rules on this page; the gaps are your reading list.
Six days out
Sit the free 53-item mock, timed. Score it as a raw count only, then for every miss write which rule would have caught it. A miss that maps to no rule means a rule is missing.
Five and four days out
Repair Applications and Integration, then Model Selection and Optimization. Between them they are almost half the paper, and both are decided by constraints rather than facts.
Three days out
One pass over the traps table and the terminology. Then the mechanical checks: your registration name matches your ID exactly, and your work email is on a recognised Claude Partner Network company domain.
Two days out
Confirm the appointment. The exam guide allows changes up to 24 hours before; pearsonvue.com and the partner FAQ say 48. Treat 48 as the deadline and you cannot be caught by either.
The day before
Nothing new. Workspace cleared, ID ready, and the time budget decided: about 40 minutes for Applications and Integration, 20 for Model Selection, and the rest in proportion.
Exam day
Read the constraint before the options. Count on every multiple-response item. Flag and move on rather than linger; the 120 minutes are fully allocated and review time is time you did not spend.
The policy facts behind those checks: the fee is $125 per attempt, and Select, Preferred and Global Premier partners pay 50 percent of it, with Global Premier partners paying nothing through 31 December 2026. Registration needs a work email on a recognised Claude Partner Network company domain. A failed attempt can be retaken after 14 days, then 30, then 90, with four attempts per exam in any rolling twelve months, each at the full fee with your tier discount applied. The credential lasts twelve months and renews on time through a free, non-proctored assessment. The same rules apply on CCAR-F, whose own cheat sheet follows the same layout for a very different blueprint. Checked against the official sources on 17 September 2026: the CCDV-F exam guide v1.0 and the partner certification FAQ.
The free mock at /practice/ccdv-f is 53 items, timed, with no account and a stated reason for every option. It reports a raw count and nothing else; no practice percentage converts to the scaled score, and only measured answers on blueprint-aligned forms inside a study account contribute to a readiness picture. Our guide to Claude certification practice exams covers what other practice material exists and how to read it.
Key takeaways
- Three domains decide the pass. Applications and Integration, Model Selection and Optimization, and Agents and Workflows are 64.6% of the paper together.
- Read the constraint before the options. Cost, latency, trust boundary or irreversibility: one stem sentence usually eliminates two options.
- Structure beats instruction. A hook, a schema, a gate or a cache placement wins over a firmer sentence in the prompt on every domain.
- Plan at skill level. Skill weights run from 1.0% to 8.6%, and the domain called Model Selection is mostly fundamentals: its namesake skill is 2.7% of the paper.
- Cap the light domains. Claude Code and debugging are about three items. Know the six rules and move on.
- A mock score is a raw count. Use it to find rules you apply inconsistently, never as a conversion to 720 or a reason to book.
Test the 33 rules on a free 53-item CCDV-F mock
Timed, no account, weighted to the official blueprint, with a reason for every option after you submit. The result is a raw count and a per-domain breakdown, which is exactly the input the final-week checklist asks for. It is not a scaled score, a pass prediction or a booking signal.
Start the free CCDV-F mockQuestions
Frequently asked
The follow-up questions people search next.
Is there an official CCDV-F cheat sheet?
No. Anthropic publishes the exam guide, which lists eight domains and 25 weighted skills, and nothing shorter. This page compresses those skills into decision rules, each paired with the trap it catches. It is derived from the guide, not endorsed by it, and it is not a list of answers.
What are the CCDV-F domain weights?
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% and Eval, Testing, and Debugging 2.6%. Three domains carry about two items in every three.
How many questions are on the CCDV-F exam and how long is it?
Fifty-three items in 120 minutes, a mix of multiple-choice and multiple-response where every multiple-response item states how many options to select. The pass mark is a scaled 720 on a 100 to 1,000 scale, and the fee is $125 per attempt before any partner-tier discount.
Is a CCDV-F quick reference enough to pass?
On its own, no. A quick reference is a compression of things you already understand, and it works as a diagnostic on practice misses rather than as reading. Used that way in the final week, after a study plan and at least one timed mock, it is the highest-yield hour you will spend.
How should I use CCDV-F decision rules under time pressure?
Read the stem for its constraint first: cost, latency, trust boundary or irreversibility. That single sentence usually eliminates two options. Then apply the domain rule to the two that remain, and check the select count before moving on. About 2.3 minutes per item is enough for exactly that.
Does a good score on the free mock mean I am ready?
No practice percentage converts to the scaled score, and the free mock reports a raw count only. Use it to find which rules you apply inconsistently, not to decide whether to book. The official result depends on a standard-setting study that no third-party paper can reproduce.
Keep reading
Related posts
Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change. Always confirm against the official exam guide before booking.