# OKH Prompt Protocol Template ### OverKill Hill P³™ — overkillhill.com/prompt-forge/ A governed prompt is not a paragraph. It is a structured operating brief. Copy this template. Fill every field. Leave nothing as "you'll know it when you see it." --- **Template version:** v1.0 **Source:** overkillhill.com/prompt-forge/#anatomy **License:** Free to use, adapt, and share. Credit appreciated, not required. --- ## THE TEN ANATOMY ELEMENTS --- ### 1. ROLE > Who or what is this model being asked to become for this task? ``` Role: [Describe the persona, expertise, perspective, or operating mode the model should adopt. Be specific. "Act as an expert" is not a role. "Act as a senior infrastructure engineer who has migrated three enterprise SaaS platforms to Kubernetes and has strong opinions about blast radius" is a role.] ``` --- ### 2. CONTEXT > What domain, prior state, constraints, and vocabulary must the model know? ``` Context: [Provide the operating environment: the project, the problem history, the vocabulary in use, the constraints that already exist, and any prior decisions that are locked. The model cannot reason well without this foundation.] ``` --- ### 3. OBJECTIVE > What is the actual job — not the surface question, the real outcome? ``` Objective: [State the real deliverable. Not what you're going to ask, but what you need to exist after the prompt runs. "Help me think through X" is not an objective. "Produce a three-option architecture decision record with tradeoffs scored against our four criteria" is an objective.] ``` --- ### 4. INPUTS > What documents, files, data, or tool access is the model drawing from? ``` Inputs: [List everything the model should reference, read, or have access to: - Attached documents - Pasted content - Tool outputs (web search, code interpreter, etc.) - Prior conversation context - External sources it should or should not consult] ``` --- ### 5. CONSTRAINTS > What must not happen? What is out of scope? What must not be invented? ``` Constraints: [Explicit out-of-bounds. This is where you prevent hallucination, scope creep, and unwanted behavior. Examples: - Do not invent statistics or citations - Do not recommend tools not already in our stack - Do not reframe the objective - Do not include legal or medical advice - Stay within the context provided; do not assume prior knowledge] ``` --- ### 6. METHOD > How should the model reason? Step-by-step? Adversarial critique? Synthesis? ``` Method: [Define the reasoning approach: - Step-by-step decomposition - Devil's advocate / adversarial critique - Comparative analysis across N options - Synthesis from multiple sources - Structured outline first, then expand - Socratic questioning before answering If no method is specified, the model picks one. Specify one.] ``` --- ### 7. OUTPUT FORMAT > What does the result look like? Fields, tables, headings, file names, artifacts? ``` Output Format: [Describe the exact shape of the output: - Markdown headers, tables, code blocks - Specific field names and data types - File naming convention if producing artifacts - Length targets (e.g., "no more than 600 words", "one page maximum") - What to omit (no preamble, no summary header, no "as an AI" disclaimers)] ``` --- ### 8. VALIDATION > How do we verify this solved the actual problem? What are the acceptance criteria? ``` Validation: [State the acceptance criteria before execution, not after: - What does a correct answer look like? - What would make you reject the output and ask again? - If the model is unsure about a field, should it flag it or fill it? - What edge case must the output handle?] ``` --- ### 9. FAILURE HANDLING > What should the model do if information is missing, unsafe, or out of scope? ``` Failure Handling: [Define graceful degradation: - If context is insufficient: ask one clarifying question or state the gap explicitly - If the request is out of scope: say so, do not invent a workaround - If the output would require unavailable information: return what is known and flag what is missing - Do not fabricate. Surface uncertainty explicitly.] ``` --- ### 10. HANDOFF > Where does the output go? Notion? Replit? GitHub? Another agent? ``` Handoff: [Define what happens after this prompt runs: - Is the output a human deliverable or agent input? - Will another model use this output as its context? - Does it feed into a Replit execution spec? - Is it being written to Notion, GitHub, or a document? Knowing the handoff forces the output format to match the destination.] ``` --- ## FILLED EXAMPLE (abbreviated) ``` Role: A technical writer who has documented five B2B SaaS products and understands the difference between feature documentation and user guidance. Context: We are documenting a Replit-hosted internal tool for a 12-person ops team. The tool has no public documentation. Current "docs" are a Slack message from March 2023. The audience is non-technical operations staff. Objective: Produce a three-section quick-start guide: What it does (one paragraph), How to access it (numbered steps), and How to get help (one paragraph). The guide should be usable without me in the room. Inputs: - Tool name and URL: [provided inline] - Existing Slack message: [pasted below] - Screenshot of the main UI: [attached] Constraints: Do not use technical jargon. Do not reference Replit or code. Do not assume the reader has ever used the tool before. Method: Outline first. Wait for approval before expanding. Output Format: Markdown. H2 headers for each of the three sections. No introduction paragraph. No footer or credits. Validation: A non-technical ops employee should be able to follow the guide with no additional explanation from me. Failure Handling: If the screenshot is insufficient to describe a step, note the gap with [NEEDS SCREENSHOT: description] rather than guessing. Handoff: Output goes into Notion. Format for pasting directly into a Notion page. ``` --- ## USAGE NOTES **Start with Objective.** If you can't clearly state what you need to exist after the prompt runs, you're not ready to write the prompt. **Constraints are not optional.** Every constraint you omit is a decision you're handing to the model. The model will make that decision. You may not like what it decides. **Handoff forces precision.** If the output is feeding another AI agent, write for that agent's input requirements, not for human readability. **Iterate with the skeleton, not around it.** If a prompt isn't working, find which of the ten elements is missing or underspecified. Don't rewrite from scratch — diagnose and fix. --- ## ABOUT PROMPT FORGE The Prompt Forge is OverKill Hill P³™'s protocol-driven promptcraft workshop. Prompts become governed AI systems, audit contracts, and Replit-ready specs. **Site:** overkillhill.com/prompt-forge/ **Methodology archive:** github.com/OKHP3/first-diagram-is-a-liar **Contact:** overkillhill.com/contact/ --- *OverKill Hill P³™ — Protocol. Process. Practice.*