To hire a Claude Code developer, test how they control the agent, not how fast it types. Look for a written CLAUDE.md, scoped permissions, hooks, subagents and MCP connections set up with care, plus automated tests and a human review of every change. Ask workflow questions, set one small paid test task on a copy of real code, and agree a rate before any production access.
A Claude Code developer is a software engineer who builds and ships production code by directing Claude Code, Anthropic's agentic coding tool, the way a lead engineer directs a junior team. The skill is not prompting. It is writing the spec, setting the rules the agent follows, splitting the work into small tasks, and reviewing and testing every change before it merges.
What does a Claude Code developer actually do?
They turn an idea into working software by steering AI coding agents and staying answerable for every line. The agent reads files, writes code, runs commands and fixes failures on its own, while the developer decides the data model, the permissions, the money logic and what ships. That division is the whole job, whether the project is a SaaS product, a CRM, a mobile app backend or an internal tool.
The distinction that matters when you hire is between someone who has asked Claude a few coding questions and someone who runs Claude Code as a controlled process every day. An agent can write a login form in a minute. It can also store passwords badly, skip a permission check, add a dependency nobody vetted, or quietly edit a file it was not asked to touch. Experience is what lets a developer design the work so those mistakes are rare and catch the rest in review. You are not hiring a prompt writer. You are hiring the judgment that sits between the agent and your production system.
So the questions you ask should be about control and verification, not about how clever the prompts are. The sections below give you the skills to check, the questions to ask, a small paid task to set, and the signals that should worry you.
Which Claude Code skills should you test?
Test the controls a developer puts around the agent, because those are what keep AI-written code safe and maintainable. Each row below maps to a documented Claude Code feature, so a candidate who uses it can explain it in plain words and show it in a repository. I checked each feature against the official documentation on 21 October 2026.
| Skill to test | What a good answer shows | Claude Code feature |
|---|---|---|
| Written project rules | Build and test commands, architecture conventions and forbidden actions written down, kept short, loaded every session | CLAUDE.md project memory |
| Limits on what the agent may do | Allow, ask and deny rules so routine commands run freely while secrets and production stay out of reach | Permissions |
| Automatic guards | Checks that run on their own, such as a formatter after an edit, tests before the turn ends, or a block on a destructive command | Hooks |
| Focused helpers | Separate agents for research or review, each with its own context and its own tool access | Subagents |
| Connections to your tools | How the agent reaches a database, tracker or docs, and with what access level, with prompt-injection risk understood | MCP servers |
| Tests and review | A check the agent can run, and a human who reads every diff before it merges | Verify its work |
The documentation is blunt about why the last row carries the most weight. The permissions page states that rules in a CLAUDE.md are context, not enforcement, so anything that must be blocked every time belongs in a PreToolUse hook. The best-practices page says Claude stops when the work looks done, and without a check it can run, "looks done" is the only signal available. A developer who cannot show the tests and the review step is selling you the agent's speed with none of its guardrails.
What should you ask a Claude Code developer in the interview?
Ask open questions about process and listen for specifics, not slogans. A strong candidate answers with file names, commands and failure stories; a weak one answers with adjectives. These six questions separate the two.
| Question | Listen for |
|---|---|
| Walk me through a real project you shipped with Claude Code. | A specific repo, the spec, how they reviewed output, and one thing the agent got wrong and how they caught it |
| What goes in a CLAUDE.md, and what does not? | Commands, architecture and "done" rules, kept under a couple of hundred lines; not a dump of the whole codebase, which the docs warn reduces adherence |
| How do you stop the agent from reading secrets or touching production? | A deny rule for the env file, least-privilege allow and ask rules, no production credentials in prompts, and bypass mode left off |
| What runs automatically on every change? | Hooks: a formatter after each edit, the test suite before the turn ends, and a block on commands that wipe data |
| Who reviews the agent's code before it merges? | A reviewer in a fresh context plus the developer reading every diff, so the agent doing the work is not the one grading it |
| How would the agent reach our database, and with what access? | An MCP server pointed at a copy with read-only or test-only access, and a note that servers fetching outside content carry injection risk |
For the full version of the process a good candidate should be able to describe, see my Claude Code workflow for client projects. If the answer to any question is "the AI handles that," ask them to show you where, in a repository, on screen.
What makes a good paid test task?
One small, real task on a copy of your code, scoped to a few hours and paid at the candidate's rate. Unpaid tests attract the wrong people and tell you little; a paid task that mirrors your actual work tells you how someone thinks under your constraints. Pick a change that touches a real part of the system but cannot break anything, such as adding a field with validation, a report, or an import of a sample file, with one clear acceptance line.
Grade the submission on how they controlled the agent, not only on whether it works. Ask for the repository, not a zip of files, so you can read the history. A short checklist to score against:
- Did they write or update a CLAUDE.md and the rules the task needed?
- Is there a failing test first, then code that makes it pass?
- Does the diff do only what the task asked, with no stray edits?
- Are secrets and real customer data absent from prompts and from the repository?
- Can they explain, line by line, a part of the diff you point at?
- Did they flag anything the agent got wrong, rather than hide it?
The last two matter most. A developer who reads back their own diff and owns its weak spots is one who will read yours. A developer who shrugs and says the agent wrote it is telling you who will be answerable when it breaks, and the answer is you.
What are the red flags when hiring a Claude Code developer?
Most of the warning signs come down to speed bought by dropping a safeguard. Treat any of the following as a reason to dig further or walk away.
| Red flag | Why it matters |
|---|---|
| Cannot show a single repository or diff | No evidence they have run Claude Code on real, reviewable work |
| Runs the agent with no review and merges whatever appears | Unread code ships, with its bugs and security holes intact |
| Pastes secrets or production data into prompts | Credentials and customer records leave your control |
| Calls themselves Anthropic "certified" or an "official partner" | There is no individual Claude Code certification; treat the claim with skepticism and ask what it means |
| "The AI tested it" with no test suite to show | Nothing proves the change works, now or after the next change |
| Fixes a fixed price to a scope nobody has written down | The price is a guess with a margin on it, and the gap becomes your problem |
| Wants to keep production credentials "to move faster" | Least privilege is ignored, so one bad command reaches live systems |
None of these is about the model being weak. They are about a person skipping the engineering that any code needs, human or AI. The tool is a public download that anyone can run; the discipline around it is what you are paying for.
How much should you pay a Claude Code developer, and how?
Pay for the engineering and the accountability, not for the agent's typing. My own rate is $55 to $65 per hour, or a fixed price for a written scope, and the Claude Code subscription I use while building is my cost, not a line on your invoice. I will not quote a marketplace rate range here, because the figures I can find sit on agency pages I cannot verify, and the brief I hold myself to allows a market figure only from a dated primary source. Ask any provider, me included, for hours by work area and the rate behind the total; quotes only compare at that level.
How you engage matters as much as the number. There are four common routes, and they are different products rather than better and worse versions of the same thing.
| Route | What you get | Watch for |
|---|---|---|
| Freelancer or contractor | One accountable person, usually hourly or fixed per scope, fast to start | Availability, and what happens if that one person is unwell |
| Agency | A team, a process and a contract a procurement department will accept | A markup for overheads, handoffs between people, and you may never meet the coder |
| Full-time hire | Long-term ownership and the deepest knowledge of your system | Weeks to recruit, and a slow, costly reversal if the fit is wrong |
| Fractional lead | Senior steering a few hours a week over your own team | Limited hours, so they direct rather than do all the building |
On freelancer versus agency in short: a freelancer gives you one accountable person at a lower rate, which suits a defined build but leaves you exposed if that person is unavailable. An agency costs more because you are paying for a team and the structure around it, which buys continuity and several specialists at once but adds handoffs and a markup. If the choice is instead between the two main agents a developer might use, Claude Code vs Codex for a SaaS build compares them.
When should you not hire a Claude Code developer at all?
When the thing you need is not what this hire provides. Hiring a developer to steer AI agents is the wrong move in several cases:
- You only need the tool set up once and you have a developer who can read the docs and learn it themselves.
- Your scope is undecided. A developer cannot make an unclear scope clear; software sets the process in code either way, so decide first.
- A contract forbids sending your code to outside AI services, unless the client agrees to it in writing.
- You expect the hire to replace a regulated system's own audits, rather than build inside them.
- The project is a throwaway prototype where the gates that protect production code cost more than the work is worth.
In the first case especially, the honest answer is often that you do not need to hire anyone. Anyone can run Claude Code; the reason to bring in a developer is that the output has to be production-safe and you cannot review it yourself. To have a project built and vetted this way, or to have Claude Code set up for your own team, see my Hire Claude Code Developer page.
What else do people ask about hiring a Claude Code developer?
How much does it cost to hire a Claude Code developer?
It depends on the route and the scope. I charge $55 to $65 per hour or a fixed price for a written scope; agencies bill a higher rate or a monthly retainer because you are buying a team. Rather than trust a headline range, send any provider your actual scope and ask for hours by work area and the rate behind the total.
Do I need a Claude Code developer, or can I use the tool myself?
You can use it yourself for prototypes and personal projects, because the tool is a public download with no special access required. You need a developer when the result has to be production-safe and you cannot review the code yourself. The agent writes fast; someone still has to decide what is correct and answer for what ships.
Is a Claude Code developer certified or partnered with Anthropic?
Anthropic does run Claude certification exams, but its certification FAQ says "Certification is available to people at Claude Partner Network organizations", and registration needs a partner company email address (read on 21 October 2026). Most independent developers therefore cannot hold one, and its absence says little about skill. Check any partner claim with Anthropic before relying on it, and judge candidates on their repositories and process. I am an independent developer and not affiliated with Anthropic.
What should a Claude Code test task include?
A small, real change on a copy of your code, with one clear acceptance line, paid at the candidate's rate. Give them the repository, not a zip, so you can read the history. Grade it on the diff, the tests they wrote, and whether they can explain a line you point at, not only on whether the feature runs.
How is hiring a Claude Code developer different from hiring any developer?
The engineering bar is the same: a spec, tests, review and a staging copy before production. The extra thing to check is how they control the agent, through written rules, scoped permissions, hooks and a review of every diff. The tool speeds up typing, not the decisions, so the judgment you are hiring for has not changed.
Last reviewed 21 October 2026. Disclosure: I sell AI-assisted development as a Claude Code developer, so I am not neutral, and this guide is written to help you vet anyone, me included. I am an independent developer, not affiliated with, certified by or endorsed by Anthropic. Claude and Claude Code are products of Anthropic; I name them only to say which tool the work uses. The tool details come from code.claude.com/docs as I read them on that date. I am also CTO at ElySpace and founder of MagizAI.
