A builder asks the question every cohort asks eventually: what do we do when we don’t really have expertise in the domain we’ve chosen?
The honest answer, unusually blunt: the domain knowledge is the hardest part: gaining all the context and engineering it. Deep-research tools help. But it is not advisable to create an agent for something you have just no clue about. There should be some secondary research involved, at minimum.
Watch the difference play out between two builders. One is an ex-SDR who encoded a hundred real cold calls into her objection patterns. The other asked an LLM to “generate common sales objections” and shipped the output. Both agents sound plausible for five minutes. Only one survives a user who has taken a real sales call, because only one contains anything the user’s own chatbot couldn’t have generated.
Why generic knowledge produces generic agents
The base model already knows the generic version of every domain. Generic sales objections, generic interview questions, generic therapy language: it has read more of each than any human alive. If your prompt adds only generic content, your agent is a thin UI over what users get from any chatbot for free, and users figure that out fast.
Value lives in the specific: this buyer persona’s actual objection sequence. This interview loop’s real rubric. This therapy modality’s actual protocol, with the actual order operations happen in. Specificity is exactly what Chapters 7, 18, and 20 kept demanding, and specificity is precisely what only research supplies.
Four surfaces of your workspace consume domain knowledge, and it is worth naming them so research has targets: scenario content (Chapter 7’s flow block), resistance and pressure patterns (Chapter 18’s directive to study real user behavior), breakthrough and evaluation criteria (the rubric mirror), and memory schemas (Chapter 11’s what-would-the-expert-note-down).
The research methods, ranked by yield
1. Your own expertise. Chapter 4’s scorecard already screened for this. If you have it, your job is extraction, which is harder than it sounds: write down your ten hardest real conversations in this domain before writing any prompt. The hard ones contain the resistance patterns; the easy ones contain nothing.
2. Expert artifacts. The highest-yield move per hour spent: get access to what practitioners actually write down. Take a peek at their meeting minutes: whatever they note down is what your agent should note down. Scorecards, call notes, session templates, intake checklists: these are pre-structured domain knowledge, begging to be encoded. An interviewer’s scorecard is your rubric block, nearly verbatim.
3. Expert interviews. One hour with a practitioner, asking one question: “walk me through your last difficult one.” Then follow-ups. The difficult cases surface the resistance patterns and teachable moments no document contains, because documents record the process and interviews record the friction.
4. Primary recordings and transcripts. Real calls, real interviews (with permission). Chapter 18’s pressure categories should be sampled from these, not invented in a brainstorm. Invented curveballs pattern-match to fiction; sampled ones land like memory.
5. Deep-research tools. The assist, properly placed: brilliant for terrain-mapping an unfamiliar domain and generating hypotheses to verify with methods one through four. Dangerous as a sole source: the output is fluent, unvalidated, and generic, which is exactly the failure mode this chapter exists to prevent.
6. UX research on your own agent. After launch, agent building becomes UX research: your transcripts are the primary source. What users actually say, where they struggle, what they ask for: Chapter 21’s loop and Part 7’s mining turn this into a permanent research program.
Encoding what you learned
Research produces notes; agents run on placed context. The placement pass, using Chapter 6 and Chapter 10’s rules:
- Behavioral patterns (how the buyer stalls, how the student deflects) go to the system prompt’s resistance and pressure blocks.
- Reference content (protocols, manuals, question banks) goes to the knowledge base, with catalog lines in the prompt.
- Judgment criteria (what “good” looks like) become the rubric block and the breakthrough conditions.
- What experts track becomes your memory instructions.
- Real dialogue snippets become few-shot examples inside prompt blocks, the single cheapest realism upgrade available. One real exchange outperforms three paragraphs of description, every time.
Example: a therapist-practice scenario, researched end to end. Two hours with a practicing therapist (“walk me through your last difficult patient”) yields four real deflection patterns, straight into Chapter 18’s pressure library. Her intake-notes template becomes the memory schema. The modality manual she trains from goes into the knowledge base with a catalog line. Her definition of a session that “worked” becomes the breakthrough criteria. Total research time: one afternoon. Result: an agent no prompt-pattern copier can replicate, because the patterns were never the moat. The encoded expertise is.
That is also the strategic point worth ending Part 6 on: a competitor can copy your prompt structure in an afternoon. Your encoded domain knowledge compounds with every iteration and every transcript, and it cannot be copied from the outside.
Built, instrumented, grounded in real expertise, iterated past its tenth version: your agent is ready for people you have never met. What happens then is the last part of this playbook.