Custom AI Agent & System Prompt Architect

Act as a staff prompt engineer shipping a production system prompt, not a personality demo. Design an agent that is constrained, testable, and host-aware. Prefer short enforceable rules over essays about being helpful.

Inputs:
- Host: [Custom GPT / Claude Project / Gemini Gem / API agent / other]
- Job of the agent: [What it must do every time]
- Users: [Who will talk to it, skill level]
- Tools / actions: [Browse, code, retrieval, APIs, or none]
- Knowledge it should use: [Files, URLs, policies it will have]
- Hard refusals: [Topics, actions, data it must not handle]
- Output shape: [Format the user should always get]
- Voice: [Tone]
- Failure I worry about: [The most likely bad behavior]

Generate:
1. Design brief (1/2 page): Job statement. What success looks like. What the agent must never do. Which failure from Inputs is the primary threat.
2. Copy-paste system prompt: A complete system prompt the user can paste. Structure:
   - Role and job (5-8 lines).
   - Input contract: what it expects, what it asks for if missing.
   - Process: numbered steps it must follow before answering.
   - Output contract: exact sections / schema.
   - Tools: when to use each, when not to. If none, say so.
   - Memory / files: what to treat as source of truth vs. conversation.
   - Guardrails: short imperative rules. Include refusals from Inputs plus hallucination, privacy, and scope rules.
   - Style: [Voice], length defaults, citation rules.
   Keep rules as bullets a model can obey. No "be creative but also accurate" mush.
3. Starter user prompt: The first message a teammate should send, with placeholders.
4. Few-shots:
   - 2 good examples: compact user -> assistant pairs that show the output contract.
   - 1 counterexample: a bad assistant reply and a one-line note of which rule it violated.
5. Eval cases (8): Table: id | user input (compressed) | should pass if | failure if. Must include: happy path, missing info, adversarial jailbreak or policy push, over-confident invention, out-of-scope request, conflicting instructions, messy real-world paste, and a tool-error or empty-retrieval case (or "no tools" equivalent).
6. Failure modes: 5 likely failures in production, signal you would see, and a prompt or process patch.
7. Host-specific notes: Custom GPT vs Claude Project vs Gemini Gem vs API: instructions length, file behavior, tool calling, what not to rely on. Give concrete setup steps for the chosen [Host] only, plus a 5-line portability note for the others.

Constraints:
- Short enforceable rules. If a rule cannot be tested with an eval case, cut or rewrite it.
- Do not invent tool APIs or file names that were not in Inputs.
- The system prompt must fit in one paste. No "part 2 of the prompt."
- Do not add a cute name, emoji persona, or secret chain-of-thought dump.
- Default: ask one clarifying question when a required input is missing, then proceed with labeled assumptions if the user says to.