← A Working Developer's Claude Code Workflow
Where does Claude Code pay off on billable work?
On the tasks that are well-specified but time-consuming: bug fixes, refactors, test repair, codebase onboarding, and scaffolding new features from an existing pattern. It pays off less on ambiguous, judgment-heavy work where the client hasn’t decided what they actually want yet, since that’s a conversation, not a coding task.
The strongest use cases
Bug fixing is a strong fit because the target is concrete: reproduce the issue, find the cause, fix it, verify. Refactoring benefits the same way, especially mechanical changes applied consistently across many files. Test repair (fixing a suite broken by a dependency bump or a refactor) is exactly the kind of tedious, well-defined work that benefits from an agent doing the grinding. Onboarding to an unfamiliar codebase goes faster with an agent that can read and summarize the whole thing before a human starts making changes. Scaffolding a new feature from an existing pattern in the codebase turns “build this the way we built that” into a fast first draft instead of a blank file.
Client reporting automation
Client work generates status updates, and Claude Code can draft a lot of that from what actually happened: commits, PR descriptions, and test results turned into a plain-language summary of what shipped. That doesn’t replace judgment about what to tell a client, but it removes the blank-page problem of writing the update from scratch every time.
The review rules that keep it safe
Nothing generated for a client project ships without a human reviewing it first, the same standard as any other engineer’s output. Security-sensitive changes (authentication, payment handling, data access) get closer review, not less, because the cost of a subtle mistake is higher. Changes that touch production data or infrastructure need explicit confirmation before they run, not just before they’re written. None of this is specific to AI-generated code, it’s the same discipline that should apply to any change on a client’s system, just enforced deliberately rather than assumed.
What not to hand it
Anything that requires a business decision, not a technical one: what a client actually wants, how to price a project, how to handle a difficult client conversation. Anything where the cost of a wrong guess is high and the correct answer depends on context Claude Code doesn’t have (a client’s specific compliance requirements, an undocumented business rule, a relationship history that shapes how something should be handled). Hand those to a person, and hand Claude Code the parts of the work that are well-specified once that person has made the call.
Talk to 84EM about bringing an AI-assisted workflow to your agency or client work if you’re figuring out where this fits.
