Build four to six finished pieces of Claude work, each showing a decision you made and the check you ran on the output. A reviewer should see the brief, the prompt approach, the result and your correction, all in non-confidential form.
The portfolio is the record around the output rather than the output on its own. Anyone can produce a summary. Fewer people can show which claim in that summary they traced back to the source, and which one they threw away.
Project briefs
What projects should I build before seeking Claude certification?
Build four: a document analysis pack, a workflow redesign with before and after evidence, a small tool or agent loop with its failure log, and a governance brief for one risky use case. Each produces evidence a reviewer can check without access to your employer's systems.
Four projects cover the four directions the role-based Claude certifications are organised around: professional knowledge work, software development, system design and enterprise delivery. Anthropic describes those four role-based certifications on its blog, and Timo Labs compares the tracks on the certifications page. Build one project per direction and you can answer a question about any track without claiming experience you do not have.
Each project carries the same five files. Write the brief before you generate anything. Keep the source material, Claude's output and the correction you made with its reason. Add one short paragraph on what the sample does not prove. That last file separates a portfolio from a screenshot folder.
Project 1
Document analysis pack
Take a public document set that nobody owns: an annual report, a regulator consultation, a published tender. Produce a structured extract and a one-page decision memo from it.
Keep the claim check list. Pick five claims in your memo and trace each one to a page and a table in the source. Show at least one claim you removed because the source did not support it.
Project 2
Workflow redesign with evidence
Pick one repeating task you already own. Time ten runs of it as it exists today, rebuild the task around Claude, then time ten runs of the new version.
Keep the run log. Record cycle time, the number of runs that needed a correction, and the correction type. A reviewer reads the failed runs first, so leave them in.
Project 3
Tool loop or small agent
Build something that calls a tool, handles the response and stops when it should. A file processor, a lookup agent or a scheduled check is enough. Scope matters less than the loop.
Keep the failure log. Record every failure mode you hit, the fix and the guard you added so it cannot return. The architect track tests this kind of design judgement directly.
Project 4
Governance brief for one risky use case
Choose a use case with real exposure: customer communication, candidate screening, anything touching money or health. Write the operating rules for it in two pages.
State what must never run without a human, name the approver by role, and define the record kept for each decision. Write it as operational guidance, and say in the document that it is not legal advice.
| Project | Judgement it shows | File the reviewer opens | Closest track |
|---|---|---|---|
| Document analysis pack | Source handling and claim checking | Decision memo plus the claim check list | CCAO-F, Associate Foundations |
| Workflow redesign | Process design and measurement | Run log across twenty runs | CCAO-F or CCAR-F |
| Tool loop or small agent | Failure handling in working code | Repository with the failure log | CCDV-F, Developer Foundations |
| Governance brief | Risk ownership and escalation | Two-page operating rule set | CCAR-P, Architect Professional |
A portfolio is not a prerequisite for any Claude exam. Anthropic sets no experience requirement for the four tracks, as the eligibility explainer sets out, and no portfolio guarantees an interview, a role or a pass. Build one because it makes your capability legible and gives you concrete material to answer questions with. The readiness guide covers the study side.
Student evidence
How can a university student build a Claude portfolio before applying for an AI role?
Convert coursework, club and volunteer work into three documented Claude projects, each with the brief, your prompt decisions, the output and the errors you caught. Student status carries no exam eligibility of its own. Registration still requires a work email on a Claude Partner Network company domain.
University work already has the shape of a portfolio project. A term assignment carries a brief, a deadline, source material and a marked outcome. What it usually lacks is a record of the decisions the student made while using Claude, and a version that can be shown to someone outside the university.
Three samples is enough for a first portfolio. Take them from three different sources so the set does not read as one module repeated. The rubric below maps what you already have to what a reviewer can open, and names what to remove before anything leaves your laptop.
| Source you already have | Sample to produce | What makes it credible | What to strip out |
|---|---|---|---|
| Term assignment or lab report | Structured analysis of the source set plus a memo on what Claude got wrong | The correction list, with the wrong output quoted next to your fix | Marks, grader comments, any unpublished dataset |
| Capstone or research project | A method note explaining where Claude was used and where you kept it out | The boundary you drew and the reason for it | Supervisor names, unpublished results, embargoed findings |
| Student club or volunteer operations | A rebuilt operating document: a rota, an intake form, an event runbook | The number of times the group actually used it after you handed it over | Member names, contact details, donor records |
| Part-time job or internship task | A generic version of the same task, rebuilt on invented inputs | The same steps and the same checks, run against data you made up | Employer name, client name, every internal figure |
| Self-directed Claude work | A small tool or agent with a failure log | The failures you kept in the log rather than the demo that worked | Nothing. This one is yours to publish in full |
Every sample passes four checks
The brief exists before the output. Write the audience, the decision the piece supports and the sources you allowed yourself, then generate.
The decision is named. State the approach you chose and the one you rejected, in a sentence each.
One claim is traced. Pick a factual claim in the output and show the source line it came from.
The limits are stated. Say what the sample does not prove, in two sentences at the end.
Student status is not exam eligibility
Anthropic sets no experience prerequisite for any of the four Claude certifications, so no candidate is filtered out on background, age or job title. Registration is the gate. Signing in requires an email on a company domain belonging to a Claude Partner Network organisation, and personal addresses do not work at sign-in. A student holding a university address sits in the same position as a professional at a non-partner employer, which is the situation the eligibility explainer describes in full.
Timo Labs offers an optional $50 mock exam with guided study cohorts. Anthropic charges the exam fee separately and directly: $99 for CCAO-F, $125 for CCDV-F and CCAR-F, and $175 for CCAR-P. You book and pay Anthropic yourself. Anthropic's exam partner hosts the exam delivery page, which is where to check current scheduling and testing conditions before you commit to a date.
Students in a department or society that wants to certify several people at once should read the sponsorship and cohorts guide. For sequencing the portfolio against a first AI role, the applied AI career guide covers what to build in what order.
Role work samples
What role-specific work samples should a business professional build with Claude?
Build one sample per function you work in: an operations SOP rewrite, a sales account brief, an HR policy summary, a finance variance narrative and a project status pack. Use invented companies and invented numbers so the sample can be shared with anyone.
A hiring manager in operations reads an SOP as an operations person. The sample is judged against the standard of the function, so it has to look like the real deliverable: same sections, same level of detail, same named person accountable at the end. A document that reads as a demonstration of Claude fails on the function's own terms, however good the prose is.
Every sample below starts with source material you invent yourself. That is the point. Inventing the ledger, the policy or the project plan forces you to define what good input looks like, and it produces a sample you can publish without clearing it with anyone.
| Function | Sample to build | Source material you invent | What the reviewer checks |
|---|---|---|---|
| Operations | SOP rewrite for one recurring process, with the exception path written out | A fictional service desk with a work order backlog and two shift patterns | Whether the exception path covers what the standard path misses |
| Sales | Account brief: current state, three questions to ask, one risk to the deal | A public-style profile you write for an invented prospect company | Whether the questions are answerable and would change how you pitch |
| HR | Employee-facing policy summary plus a manager FAQ | An invented leave or conduct policy that you draft in full first | Whether the summary shifts meaning anywhere against the source policy |
| Finance | Variance narrative explaining the movement in one budget line | A small invented monthly ledger you build in a spreadsheet | Whether every figure in the narrative reconciles to the sheet |
| Project work | Status pack: progress, decisions needed, risks with owners and dates | An invented three-workstream plan with two dependencies between them | Whether each risk carries a named owner and a date |
Illustration: an operations SOP rewrite
Northwind Facilities is fictional and every figure here is invented. The brief: rewrite the work order intake procedure so a new coordinator can run it in week one, with a backlog of 42 open orders and two shift patterns to cover.
What the sample contains. The original procedure as a four-step list, the rewritten procedure with an exception path, a correction log, and a limits note.
The decision recorded. The first draft merged the intake and triage steps into one. The coordinator would then own a judgement call with no rule behind it, so the steps were separated and a triage rule was written for each priority level.
The error caught. The draft assumed a contractor is always available. The exception path now routes an unavailable-contractor case to the shift supervisor with a two-hour response window, and the correction log quotes the original sentence next to the replacement.
The limit stated. The procedure was never run in a live facility. It shows document design and exception handling, and it shows nothing about adoption.
Use invented companies and invented numbers throughout. A sample built from employer data cannot be shared, and a reviewer cannot verify it either, so it fails twice. Strip client names, internal metrics and anything covered by an agreement before a file leaves your machine. Where a process detail is genuinely distinctive to your employer, change it rather than remove it, and say in the limits note that the detail was altered.
Non-technical professionals building toward CCAO-F will find the track comparison in the non-technical certification guide. Document-heavy roles should pair the analysis sample with the document workflow guide. Hiring managers assessing this kind of evidence can start at the employers page.
Packaging
Package the portfolio so a reviewer can check it in ten minutes
A reviewer gives a portfolio a few minutes before deciding whether to open anything. Structure carries that decision. Use one index page listing each sample in three lines: the brief, the decision you made, the link. Give every sample the same folder shape, so the second sample takes a fraction of the attention the first one did.
- Name files after the deliverable. A file called quarterly-status-pack tells a reviewer what it is; a file called claude-output-3 tells them nothing.
- Keep the folder shape identical. Brief, source, output, correction log, limits, in that order, in every sample.
- Lead with the correction log in conversation. The error you caught is the part that shows judgement, and it is the part most candidates hide.
- Cap the set at six. Six checked samples read better than fifteen unchecked ones, and you can speak to all six from memory.
Answer questions about each sample in two sentences: what you decided, and what you rejected. The questions page covers what candidates ask most often about the process around them.
Related Timo Labs guides
Sources
- Anthropic: four role-based Claude certifications, checked 2026-08-13.
- Anthropic's exam partner: exam delivery page, checked 2026-08-13.
- Timo Labs: Claude Certified Architect, checked 2026-08-13.
- Timo Labs: who can take the Claude certification exam, checked 2026-08-13.
- Timo Labs: Claude certification for non-technical professionals, checked 2026-08-13.
Anthropic runs the Claude Certification programme, sets the exam fees and awards the credential. Amotion AI is a member of the Claude Partner Network. Timo Labs supports preparation with an optional mock exam and guided study cohorts. Exam fees are Anthropic's published fees, paid at registration. The governance guidance on this page is operational and does not replace legal advice.
Use the portfolio alongside certification preparation
An optional $50 mock exam with guided study cohorts. Anthropic charges the exam fee separately, and you register and pay directly.