# The agent prompt anatomy — field checklist Companion artifact to curriculum module 2.01 (Generative AI agent design) at https://lyceumagents.com/curriculum/agent-design/ This is the anatomy the production agents in the melchizedek repo actually use: an instruction subdivided into named blocks, each with one job. In sophisticated agents the blocks are literal XML-tagged sections (``, ``, ``, `` — see patient_advocate.yaml); in simpler agents the same anatomy appears as markdown headers, PHASE blocks, or all-caps MANDATE lines. The framework treats the instruction as one string — the subdivision is an authoring discipline, not a syntax. Blocks follow capabilities: an agent earns a new doctrine block the moment it gains a new power. --- ## Before any block: the purpose - [ ] Write the one-sentence purpose the agent serves. If you cannot, you are configuring a general assistant, not building an instrument. Every block below enforces this sentence. - [ ] Strip and anonymize the input. PII, credentials, and contextual metadata come out BEFORE the model sees anything. ## Identity — `` - [ ] Name the role and its expertise ("a patient and exacting private teacher", "a specialized Patient Advocate and Medical Co-Pilot"). - [ ] Name the USER and their situation — who is being served, in what state ("navigating a medical situation… needs an expert, comforting, yet highly direct peer"). - [ ] State the purpose and its edge: what the agent is for, and what requests get redirected because they fall outside it. ## Doctrines — one per capability ``, ``, ``, `` — each governs how ONE capability may be used: - [ ] Name the ground truth ("the pasted material", "this patient's record") and require every substantive claim to anchor to it. - [ ] Legislate the silence: when the source holds no answer, the agent says so — it never fills the gap ("never invent a remembered fact"). - [ ] Require provenance: where a fact came from, with date and source, cited when used. - [ ] For each tool or subagent: when it MUST be consulted, and what may never be asserted without it ("never state drug dosing claims from memory — verify through MedScribe"). ## Communication laws — `` - [ ] Lead with substance: the first sentence carries the answer or the critical warning. Ban filler openings by name. - [ ] Make every law countable: "exactly one follow-up question", "maximum 90 words", "define every term inline at first use". A law you can't check against a transcript is a wish. - [ ] Fix the layout contract: headers, bullets, blockquotes — decided here, not improvised per run. ## Workflow — `` / PHASE blocks (Appears when the agent runs multi-step work — see image_production.yaml and the financial agents.) - [ ] Number the phases; give each an exit condition. - [ ] Place the confirmation gate: which phase may not begin without explicit human approval ("NEVER call generate_image in Phase 1"). - [ ] State the delegation order: who must be consulted, in what sequence, with obligations marked MUST. ## Boundaries — `` / `` - [ ] Behaviors, not disclaimers: rules that change what the agent DOES at the edge, not boilerplate appended to outputs. - [ ] Write the overrides first: the conditions under which everything else yields (emergency instructions come before explanations). - [ ] Divide the labor: what the agent prepares vs. what stays with the human professional ("you prepare, they prescribe"). - [ ] Legislate uncertainty: uncertain answers are ranges with the unknown named, delivered plainly — "calm and honest beats soothing and vague". ## Outside the instruction (the config is part of the prompt) - [ ] Each subagent's `description` is its delegation API — write it like a function signature: when to call, what to pass. - [ ] Temperature follows the job: warm where judgment synthesizes, cool where facts are retrieved (0.7 mentor / 0.2 researcher). - [ ] Output schemas make quality machine-checkable — a parsed integer can gate a loop; prose approval cannot. ## After the output — verification stage - [ ] Fact-check every claim, citation, and figure. Fluency is not evidence. - [ ] Never let the generator grade itself — a checker is separate from the maker. - [ ] You own the output. The model multiplied your speed, not your responsibility. --- ## When output fails: the five debug variables Adjust these before assuming the model is broken (worked fully in module 2.02): 1. **Clarity** — remove ambiguous jargon and tangled syntax; speak literally to a literal machine. 2. **Precision** — enforce hard boundaries: exact counts, tone, scope. 3. **Context** — supply the background the task silently assumed. 4. **Instruction** — swap weak verbs for cognitive ones (synthesize, contrast, extract). 5. **Format** — demand the exact output structure: table, JSON, list.