AiImported version
Prompt Jason
/Users/fabien/Desktop/AI/PROMPTS/-ai/-prompting/optimization/-prompt-jason.odt
Folder-ai/-prompting/optimization
Versions0
Imported2026-06-28
StatusActive
## JSON Prompting (Mini Guide) ### Step 1. Define one outcome and the success test Give the model a single job and how you will judge it. State the user, the problem, the format, and the success criteria. Keep it tight. Copy lines: - Goal: [one clear outcome]. - Audience: [who it is for]. - Success criteria: [exact items you expect]. - Length limit: [words or bullets]. If you cannot write this in 3 lines, split it into another prompt. ### Step 2. Structure your inputs into labeled fields Feed context as labeled chunks so the model knows what is what. Use short labels and simple values. Avoid paragraphs. Copy labels you can reuse: - Context: [what matters]. - Sources: [titles or URLs]. - Constraints: [no fluff, neutral tone, cite if used]. - Keys required: [the fields you want back]. - Don'ts: [what to avoid]. This turns your prompt into a form the model can fill rather than guess. ### Step 3. Force a JSON response, explicitly Tell the model the exact keys you want and to return only valid JSON. No prose, no preamble. Say what to do if information is missing. Copy lines: - Return only valid JSON with these keys: [key_1, key_2, key_3]. - Use arrays for lists. Use strings for text. - If a field is unknown, set it to null and ask up to 3 clarifying questions first. - Do not add extra keys. Do not add commentary. This one step is what makes outputs consistent and automation-ready. ### Step 4. Add rules and a self-check before finalizing Bake in guardrails and a final quality pass. You want the model to verify itself, then revise. Copy lines: - Constraints: [word cap], [style], [objectivity], [no speculation]. - Fact check: verify claims with 2 credible sources. Include a sources list. - Logic check: flag gaps, contradictions, or missing steps. - Revise: improve the draft based on the checks, then return the final JSON only. This improves accuracy and reduces cleanup. ### Step 5. Template, version, and iterate Save your best prompt as a template. Name versions, change one thing at a time, and keep a small library of reusable lines. Keep these snippets handy: - Clarifier: ask exactly 3 questions if required info is missing, then proceed. - Format enforcer: return only the requested JSON keys, nothing else.