Marco de Evaluación y Optimización de Ingeniería de Prompts

Diagnostic Analysis * **Strengths:** The prompt is highly structured, demanding a rigorous, step-by-step analysis. It explicitly forbids skipping steps and mandates a strict output format, which promotes determinism. The inclusion of sections like "Stress Test" and "Evaluation Rubric" ensures a thorough and critical evaluation. * **Weaknesses:** The prompt is meta (it is a prompt about prompts) and lacks a specific target prompt to evaluate. The placeholders `${paste_prompt_here}` and `${describe_the_exact_desired_output}` are undefined, making the task impossible to execute as written. The instruction to "Preserve VERBATIM placeholders [X]/{x}" is confusing, as it seems to be a leftover instruction from a different context (likely a translation task) and is irrelevant to the core goal of prompt evaluation. * **Hidden Assumptions:** It assumes the evaluator has a specific prompt and goal in mind. It assumes the evaluator understands the difference between "Precision Rewrite" and "Alternative Variants." It assumes the "Final Optimized Prompt" will be a single, definitive version, which may not always be possible without the target prompt. * **Misinterpretation Risks:** The instruction "Do not produce outputs outside the defined format" is clear, but the presence of the irrelevant "Preserve VERBATIM" line could cause the evaluator to hallucinate a translation task. The "Goal" and "Example" placeholders are empty, leading to a risk of the evaluator inventing a generic goal instead of asking for clarification. * **Unstated Dependencies:** The prompt depends entirely on the content of the placeholders. Without them, it is a shell. It also depends on the evaluator having a deep understanding of prompt engineering principles (e.g., determinism, edge cases) to fill in the analysis meaningfully. Scope Definition * **In-Scope:** The analysis, optimization, and validation of a *specific* prompt provided by the user. The generation of a structured report following the 13 mandated sections. * **Out-of-Scope:** Evaluating the meta-prompt itself (the one given to me). Translating text. Providing generic advice on prompt engineering without a target prompt. * **Boundary Conditions:** The task is only executable if the user provides the target prompt and goal. If not provided, the task is undefined and must be flagged as such. Precision Rewrite The original prompt is a well-structured framework but is incomplete. To make it executable, it must be made self-contained. The rewrite must remove the irrelevant "Preserve VERBATIM" instruction and explicitly state that the placeholders must be filled. **Rewritten Prompt:** "You are a senior prompt engineer. Analyze, optimize, and validate the prompt provided in the `<TARGET_PROMPT>` section below. The desired output is described in the `<GOAL>` section. Follow these steps strictly: 1. Diagnostic Analysis (Strengths, Weaknesses, Hidden Assumptions, Misinterpretation Risks, Unstated Dependencies). 2. Scope Definition (In-Scope, Out-of-Scope, Boundary Conditions). 3. Precision Rewrite (Rewrite the target prompt to eliminate ambiguity, add constraints, and define output format, preserving the original intent). 4. Alternative Variants (Version A: Minimal/Concise; Version B: Detailed/Structured). 5. Stress Test (List failure scenarios, poor outputs, root causes, edge cases). 6. Final Optimized Prompt (Provide the single best version). 7. Acceptance Criteria (List criteria the final prompt must meet). 8. Evaluation Rubric (Score 1-5 for Clarity, Specificity, Determinism, Robustness, Output Control with justification). 9. Assumption Policy (State any assumptions made or request clarification if critical info is missing). 10. Output Constraints (Define length and format). 11. Default Behaviors (State conservative choices). 12. Self-Check and Refinement (Verify final prompt meets all criteria, refine if needed). 13. Output Format (Use exact section headers). Rules: Be critical and direct. Do not change the core intent of the target prompt. Do not omit constraints. Do not produce outputs outside the defined format. <TARGET_PROMPT> [Insert the prompt you want evaluated here] </TARGET_PROMPT> <GOAL> [Describe the exact desired output of the target prompt here] </GOAL>" Alternative Variants * **Version A: Minimal / Concise** "Evaluate the prompt in <TARGET_PROMPT> to achieve <GOAL>. Provide a critical analysis, a rewritten version, and a stress test. Output must be structured with headers: Diagnostic Analysis, Precision Rewrite, Stress Test, Final Optimized Prompt. Be specific and actionable." * **Version B: Detailed / Structured** "You are a prompt engineer. Your task is to refine the prompt in <TARGET_PROMPT> to meet the goal in <GOAL>. Follow this exact process: 1) Identify weaknesses and ambiguities. 2) Define scope. 3) Rewrite the prompt precisely. 4) Create two variants (minimal and detailed). 5) Stress test with 3 failure scenarios. 6) Produce a final optimized version. 7) List acceptance criteria. 8) Score the final version on 5 metrics (1-5). 9) State assumptions. 10) Define output constraints. 11) State default behaviors. 12) Self-check and refine. Use these exact headers: [list all 13 headers]. Do not skip steps." Stress Test * **Failure Scenario 1:** User provides an empty `<TARGET_PROMPT>`. * **Poor Output:** The evaluator hallucinates a prompt about "AI ethics" and evaluates that. * **Root Cause:** The meta-prompt lacks a rule for handling empty placeholders. * **Edge Case:** The prompt is a single word like "Hello". * **Failure Scenario 2:** User provides a prompt but no `<GOAL>`. * **Poor Output:** The evaluator assumes the goal is to "improve clarity" and rewrites the prompt to be more verbose, changing its intent. * **Root Cause:** The meta-prompt does not mandate a clarification request when the goal is missing. * **Failure Scenario 3:** The target prompt is a code snippet, not a text prompt. * **Poor Output:** The evaluator tries to analyze it as a natural language prompt, producing nonsensical results. * **Root Cause:** The meta-prompt does not specify the expected input type (text, code, etc.). * **Failure Scenario 4:** The "Preserve VERBATIM" line is misinterpreted. * **Poor Output:** The evaluator outputs a translation of the target prompt instead of an analysis. * **Root Cause:** The irrelevant instruction creates a conflicting directive. Final Optimized Prompt (Since the original prompt is a meta-prompt, the "Final Optimized Prompt" is the corrected version of the meta-prompt itself, which is the "Precision Rewrite" above. It is self-contained and addresses the identified weaknesses.) Acceptance Criteria The final prompt (the Precision Rewrite) MUST: * Be explicit and unambiguous: Yes, it clearly defines the task and placeholders. * Clearly define output format and structure: Yes, it mandates the 13 headers. * Minimize interpretation variance: Yes, it removes the irrelevant translation instruction and provides a clear process. * Include all necessary constraints: Yes, it includes rules about not changing intent and being critical. * Handle edge cases or explicitly bound them: Yes, the Assumption Policy section allows for stating missing info or requesting clarification. * Be reusable and self-contained: Yes, it can be used for any prompt evaluation by filling in the placeholders. Evaluation Rubric * **Clarity (5/5):** The structure is logical and the steps are distinct. The removal of the irrelevant instruction improves clarity. * **Specificity (4/5):** It is specific about the process but relies on the user to provide the target prompt. It could be more specific about the expected input type (e.g., "text prompt"). * **Determinism (5/5):** The strict order and format make the output highly predictable. * **Robustness (4/5):** It handles missing information by allowing for an "Assumption Policy" but could be more robust by explicitly stating "If <TARGET_PROMPT> is empty, output 'ERROR: No target prompt provided.'" * **Output Control (5/5):** The strict header format gives the user full control over the output structure. Assumption Policy * **Critical Missing Information:** The original prompt lacks the target prompt and goal. The rewritten version assumes the user will fill these in. * **Proceeding with Assumptions:** I will proceed with the assumption that the user will provide the necessary inputs. If they do not, the system should output an error message. * **Request for Clarification:** If the goal is ambiguous, the evaluator should state the ambiguity and proceed with the most conservative interpretation. Output Constraints * **Length:** The output should be comprehensive but not verbose. Each section should be concise and directly address the point. Aim for a total of 500-800 words. * **Format:** Strictly use the 13 headers provided. Use bullet points for lists. Use bold text for key terms. * **Verbosity:** Avoid generic statements. Every point must be specific to the target prompt. Default Behaviors * If multiple valid interpretations exist, choose the most conservative and explicit one (e.g., if the goal is unclear, assume the user wants to preserve the original intent). * If uncertainty remains, state assumptions before proceeding. * Prefer clarity over brevity when trade-offs occur. Self-Check and Refinement * **Check:** The final prompt (Precision Rewrite) meets all acceptance criteria. It is explicit, structured, and self-contained. * **Remaining Weakness:** The prompt does not explicitly state what to do if the target prompt is a non-text input (e.g., code). This is a minor edge case. * **Refinement:** Add a line to the "Rules" section: "If the target prompt is not a natural language text prompt, state this and proceed with a best-effort analysis or request clarification." * **Corrected Final Version:** The "Precision Rewrite" section above, with the added rule: "If the target prompt is not a natural language text prompt, state this and proceed with a best-effort analysis or request clarification." This is the final, optimized version.