How to front-load your professional knowledge into an AI agent session — and why it matters more than knowing how to prompt.
Anthropic's June 2026 analysis of 400,000 Claude Code sessions found that management, legal, and sales roles completed AI agent tasks at nearly the same verified rate as software engineers. The differentiator was domain expertise — how well someone understood the problem — not coding skill. This post gives you the briefing template that transfers that expertise into every agent session before the task instruction is written.
Most people who start using AI agents assume technical fluency is the decisive variable. The people who get the most out of these tools, they imagine, are the ones who know how to write prompts, chain instructions, and configure workflows. The data does not support this.
This pattern is drawn from Anthropic's June 2026 research paper "Agentic coding and persistent returns to expertise", a privacy-preserving analysis of roughly 400,000 Claude Code sessions from about 235,000 users between October 2025 and April 2026. The headline finding: on verified task completion, non-software occupations reached a 29% success rate and software engineers reached 34% — a five-point gap that is smaller than the gap between novices and experts within any single occupation. Management roles, on average, slightly outperformed software engineers. The fastest-growing non-software groups using Claude Code were management, sales, and legal.
The study found that in a typical session, the human makes most of the planning decisions — what to do, what constraints apply, what the output should look like — and the AI makes most of the execution decisions. The greater the domain expertise the person brought to the planning side, the more work the AI completed per instruction.
Novice users reached verified success only 15% of the time. Intermediate and expert users reached 28–33%. The gap between novice and expert was far larger than the gap between occupations.
If you are a lawyer, an analyst, a sales manager, or a product lead, your professional knowledge is not incidental to the AI agent session. It is the primary ingredient. The bottleneck is not the prompt — it is getting your domain knowledge into the session in a form the agent can use.
The domain expertise brief is a short preamble you write before the task instruction in every agent session. It takes about five minutes. It does four things: states who you are and what you specifically know about this domain; defines what done well looks like in concrete terms; names the failure modes you have actually seen; and sets a hard constraint on evidence.
Copy this template and fill in the brackets before you write a single task instruction:
Context: [Your role and what you specifically know about this domain — not just your job title, but the distinctions that matter in your field: terminology, edge cases, what good looks like versus what plausible-but-wrong looks like]
Task: [One specific instruction — not a topic, a concrete deliverable]
What done well looks like: [Observable signal — not "high quality" but specific criteria you would use to accept or reject the output]
What done badly looks like: [Two or three failure modes you have actually seen — invented figures, wrong terminology, wrong audience framing, missing the key issue]
Constraints: [What the agent must not do — invent data, draw conclusions beyond the evidence, change the scope, reframe the audience]
Evidence requirement: Before closing out, list every factual claim you made and where each one came from.Here is the same template filled in by a sales operations manager producing a pipeline health summary:
Context: I am a Sales Operations Manager covering EMEA mid-market accounts. I understand this pipeline deeply: how deals move through stages, which stage definitions matter (we use "Proof of Concept", not "Demo"), and what at-risk means in our context — stuck in Proposal for more than 14 days, no champion identified, or single-threaded with no executive sponsor.
Task: Analyse the attached CRM export and produce a pipeline health summary for the VP of Sales.
What done well looks like: Flags at-risk deals by name, shows quarter-on-quarter slippage by stage, readable in two minutes, can go to the VP without editing.
What done badly looks like: Generic summaries without deal names, invented estimates when data is missing, confusion about our stage names.
Constraints: Use only numbers from the attached CSV. If a field is blank, say so explicitly — do not estimate or interpolate.
Evidence requirement: At the end, list every figure cited and the row it came from.The context block is where domain expertise enters the session. A vague "help me with a pipeline report" requires the agent to guess at context it cannot see. An explicit statement of your role, your terminology, and the distinctions that matter in your field removes that ambiguity before the task instruction is written. The Anthropic study found that domain expertise scales with how much work the agent completes per instruction — this is the mechanism.
Done well / done badly replaces the generic quality standard "make it good" with a picture the agent can recognise. Most people skip this. It is also the part that most directly prevents the failure mode where the output looks polished but fails on criteria the agent never knew to apply — because you never stated them.
Constraints keep planning authority with you. The study found that humans make planning decisions and agents make execution decisions. Constraints are the boundary. Saying "do not estimate if data is missing" is not defensive over-engineering — it is the difference between a report you can send and one that causes a problem you have to fix later.
The evidence requirement puts verification inside the session rather than outside it. It takes the agent about 30 seconds. It is not a guarantee against hallucination, but it is the fastest way to catch a fabricated figure before the output leaves your hands.
"The greater domain expertise a person brings to a session, the more work Claude does per instruction."
The brief breaks in predictable ways.
The most common: the context block describes your job title but not your knowledge. "I am a sales manager" tells the agent very little. "I understand which stage definitions matter and what at-risk looks like in our context" gives it something to act on. Job titles are shorthand; domain knowledge is the actual input.
The second: the evidence requirement is omitted because it feels formal. In practice it takes 30 seconds and it is the fastest way to surface a fabricated figure before it becomes someone else's problem.
The third: the brief is written once and reused verbatim across different task types. Constraints written for a pipeline report will actively mislead an agent producing a competitive landscape summary. The brief should be task-specific, not a general system prompt saved once and reused indefinitely.
The Anthropic study's finding is not that non-coders are as good at writing code as engineers. It is that domain expertise transfers into AI agent sessions — and the people who bring it explicitly, in the planning phase, get more useful work out of every session. The brief is the transfer mechanism. Write it before the task instruction, not after. Your professional knowledge is the asset; the brief is how you put it to work.