--- title: "What Is Role Engineering? Manage Claude Like Staff" url: "https://goremy.ai/blog/articles/role-engineering-managing-claude-like-an-employee" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-25T08:40:07.146Z" updated: "2026-10-25T08:40:07.146Z" topic: AI Agents & Automation description: Role engineering replaces prompting with delegation. Learn SMART goals, knowledge bases, gold-standard examples and SOPs for long-running Claude work. --- # What Is Role Engineering? Manage Claude Like Staff > Role engineering replaces prompting with delegation. Learn SMART goals, knowledge bases, gold-standard examples and SOPs for long-running Claude work. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-25 ## What is role engineering? Role engineering is the practice of treating an AI model as a persistent staff member instead of a tool you phrase requests to. You give Claude a measurable goal, the background it needs, and an example of good work. Then you let it run, review the output, and turn your feedback into reusable procedures. The skill shifts from wording prompts to delegating, judging and training. ## TL;DR - Role engineering moves your effort from **phrasing the perfect prompt** to defining the job, supplying context, and reviewing results. - Anthropic's own guidance frames Claude as a **brilliant but new employee** who lacks context on your norms and workflows. - Of the SMART criteria, **measurable (verifiable) goals** matter most for agents, because an agent that can't check its work can't know when it's done. - You will likely have to **do the task yourself once** to produce a gold-standard example the agent can compare against. - Run the task five or ten times, collect your critiques, and loop until they're fixed. The result is a **written SOP** you can hand to any future agent. - Keep your task list so you can **re-benchmark each new model** and compare quality and cost. ## Why is prompt engineering giving way to delegation? Models have become capable generalists. Phrasing matters less when the model can infer intent from a rough outline, the same way a competent employee can. The shift also shows up in how long agents work unattended. Boris Cherny, who created Claude Code, and other Anthropic staff have said they hand Claude a prompt and let it run for weeks in the background, checking in on progress and steering it toward a final result. The creator of the source video also cites a claim that AI now writes over 80% of the code merged inside Anthropic, through very long sessions. Treat that figure as a reported claim, not an audited statistic. Two examples show the pattern: - **The Swift rewrite.** Cherny wanted the Claude desktop app rewritten in Swift. The instruction, as relayed, was to rewrite the Electron app in Swift, run the Electron app in a Mac virtual interface, screenshot it, compare it pixel by pixel to the Swift version, and not stop until done. It ran for over two weeks. - **The C compiler.** Anthropic researcher Nick Carlini had 16 Claude agents work together to build a C compiler. The instruction was essentially to keep going until it was perfect. That also ran for about two weeks. Notice what both prompts share. They are short, but each has a clear finish line the agent can check. The Swift task had a visual comparison. A compiler has a definable standard of correctness. ## How do you delegate to Claude like a staff member? Anthropic's guidance says to treat Claude like a brilliant but new employee who lacks context on your norms and workflows. In practice that means four inputs: 1. A goal. 2. The background behind it: how it came about and why you want it. 3. An example of high-quality work. 4. Any mistakes you want to avoid. The rest of this article walks through each. ## How do SMART goals apply to AI agents? SMART is the long-standing manager's framework for delegation: specific, measurable, attainable, relevant, time-bound. A task should have all five. For agents, measurable carries the most weight. If a goal is verifiable, the agent can check its own work and know whether it hit the target. If it isn't, it has nothing to check against. | Vague goal | Measurable goal | |---|---| | Make my website better | Get the homepage to load in under 2 seconds on an iPhone display | The second version gives the agent a test to run, a result to read, and a reason to keep iterating or stop. The Swift rewrite works the same way: the screenshot comparison is the measurement. ## What context does Claude need, and where does it live? An employee builds up context over months. An agent starts each session with none unless you supply it. The older approach, context engineering, meant assembling that background by hand for every task. The alternative is a persistent knowledge base the agent can search. The source creator describes one holding over 50,000 of his own records: emails, video transcripts, past AI conversations across several models, GitHub commit logs and repos. The point is that you don't rebuild context each time. The agent can look up your motivations, past decisions and preferences itself. You also need connectors to the software services the agent must touch. The creator notes that newer agent products handle all of that authentication plumbing in the background. The principle holds regardless of tool: an employee with no access to your systems can't do the job. ## Why do you need a gold-standard example? Telling a model to "do a great job" doesn't work, because it has no definition of great. It needs something to compare against. That means doing the task yourself at least once. The creator calls this probably the most valuable time you'll spend, since it produces the reference everything else is measured against. Once you have it, you hand it over with an instruction like "build it just like this." Include examples of mistakes to avoid if you have them. ## How do you turn feedback into SOPs? Don't judge Claude on a single run. Have it do the task five or ten times, then review the batch the way you'd review a staff member's work. Then: 1. Read all the outputs and note the typical quality. 2. Write specific critiques per run. For example, in run A something was done in a way you disliked, in run B a required element was missing, in run C a step was wrong. 3. Compile these into one list of changes. 4. Give the list to Claude and loop until every item is fixed. What you end up with is a written standard operating procedure. It lives in a format you can give to any agent later and have it reproduce the work. This costs tokens and time, and the creator is frank that it takes a large share of his working day. The trade is that you are training the process as much as the agent. ## How do you re-benchmark when a new model ships? Keep the list of tasks the agent routinely performs, along with your gold-standard examples and SOPs. When a new model arrives, run the same tasks on both models and compare. You get a direct read on quality improvement, and you can compare cost per task as well. The source speculates that jumps between model generations could be large on many tasks, with error rates falling sharply. That is a prediction. The benchmark list is how you test it against your own work instead of taking anyone's word. ## How much does it cost to run Claude like an employee? Long-running agents burn tokens. The source creator says he spent around $2,500 in tokens in a single day, which extrapolates to roughly $50,000 a month over 20 working days. He calls himself an extreme case, and your numbers will differ. The point is that the bill can climb quickly at API pricing, so audit it the way you'd audit headcount. Cost levers mentioned: - **Match model to job.** Use Haiku for low-level, straightforward tasks, Sonnet for run-of-the-mill knowledge work, and Opus for high-value work. - **Protect prompt caching.** The creator says a repeat prompt within a window can cost only 5 to 10% of the normal rate. Long gaps between sessions forfeit that discount, so compress related work into tighter sessions. Check Anthropic's current pricing documentation for exact terms. - **Batch non-urgent work.** If a task has a turnaround of a day or more, batch APIs let you submit a large request, such as overnight, and collect results later. - **Don't over-hire.** The creator suggests one agent covers most tasks for most people. You can embed that agent in a team communication channel so everyone prompts the same one. ## Is role engineering worth learning? If your work involves tasks that are repeatable and checkable, probably yes. The method rewards people who can define outcomes, judge output and write down what "good" looks like. Those are management skills, and they transfer across models. It is less suited to tasks you can't verify. If you can't state a measurable outcome, the agent can't either, and you'll be stuck reviewing everything by hand. Start with one task that has a clear test, do it once yourself, and build from there. ## Frequently Asked Questions ### Is prompt engineering obsolete? Not entirely, but its role has shrunk. Clear instructions still matter. But with capable models, the payoff has moved to goals, context, examples and review. You no longer need elaborate phrasing tricks for most tasks. ### What is the most important part of a SMART goal for an AI agent? Measurable. A verifiable goal lets the agent test its own work and know when to stop. "Load the homepage in under 2 seconds on an iPhone display" gives it something to check. "Make the website better" does not. ### Do I really have to do the task myself first? The source strongly recommends it. Your own completed work becomes the gold-standard example the agent compares against. Without it, the model has no concrete definition of what good looks like for your situation. ### How many times should I run a task before writing an SOP? The source suggests five or ten runs. That gives you enough outputs to spot recurring problems, write specific feedback, and loop until those issues are fixed. Expect the process to cost extra tokens and time up front. ### How do I keep costs under control? Match model size to task difficulty, keep sessions tight to benefit from prompt caching, and batch work that isn't urgent. Review spend regularly, and avoid running many agents when one handles most of the load.