The most common reason a prompt fails is not bad phrasing — it’s missing information. A prompt that asks a model to “write a summary” without specifying the output length, the audience’s technical level, or what counts as a good summary will produce wildly inconsistent results. Prompt Master solves this by decomposing every request into nine structured intent dimensions before writing a single word of the actual prompt. Three dimensions are extracted unconditionally. Six are extracted only when relevant. And if critical dimensions are still missing after extraction, Prompt Master asks — but never more than three questions.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/nidhinjs/prompt-master/llms.txt
Use this file to discover all available pages before exploring further.
The 9 Intent Dimensions
| Dimension | What to Extract | Critical? |
|---|---|---|
| Task | Specific action — convert vague verbs to precise operations | Always |
| Target tool | Which AI system receives this prompt | Always |
| Output format | Shape, length, structure, filetype of the result | Always |
| Constraints | What MUST and MUST NOT happen, scope boundaries | If complex |
| Input | What the user is providing alongside the prompt | If applicable |
| Context | Domain, project state, prior decisions from this session | If session has history |
| Audience | Who reads the output, their technical level | If user-facing |
| Success criteria | How to know the prompt worked — binary where possible | If task is complex |
| Examples | Desired input/output pairs for pattern lock | If format-critical |
Always-Extracted Dimensions
Three dimensions are extracted on every single request. If any of the three cannot be resolved from your input or the session context, the pipeline pauses before writing the prompt. Task is the most frequently underspecified dimension. Natural language requests are full of vague verbs — improve, enhance, fix, make it better. During extraction, every vague verb is converted to a precise operation: improve becomes restructure the argument sequence; fix becomes correct the type mismatch in the function signature at line 42; make it better becomes increase output specificity by adding concrete examples for each claim. This conversion is what allows the resulting prompt to give the model an unambiguous instruction rather than a hope. Target tool determines the entire structural approach of the prompt. Claude, GPT-4o, Gemini, Midjourney, Sora, Runway, and the reasoning-native models (o3, o4-mini, DeepSeek-R1, Qwen3 thinking mode) each have different optimal prompt architectures, context window behaviors, and instruction-following patterns. Routing the wrong structure to the wrong tool is the single most common cause of first-attempt failure. Output format covers shape, length, structure, and filetype. A prompt that does not specify output format leaves the model free to invent its own — and models frequently invent formats that are technically correct but practically useless. Explicit output format instructions (Respond as a JSON object with keys: title, summary, tags) eliminate an entire category of post-processing work.
Conditionally-Extracted Dimensions
The remaining six dimensions are extracted when they are present or relevant to the task. Extracting them when they are not relevant adds noise without signal — a creative writing prompt rarely needs success criteria; a data pipeline prompt rarely needs audience calibration. Constraints — Extracted when the task has meaningful scope boundaries or hard prohibitions. Constraints are always placed in the first 30% of the final prompt, before any instructional content, because the most important thing the model needs to know is what it must not do. Input — Extracted when you are providing material alongside the prompt (a document, a dataset, a codebase, an image). Knowing that input exists changes how the prompt is structured — it needs to reference and handle the input explicitly rather than asking the model to generate from scratch. Context — Extracted from session history when the current request builds on prior decisions. If earlier in the conversation you established a tech stack, locked an architecture pattern, or tried an approach that failed, that context is captured in a Memory Block and prepended to the prompt. Without this, the model starts fresh and ignores decisions that are supposed to be settled. Audience — Extracted when the prompt’s output will be read by a specific person or group. A technical explainer for a senior engineer is structured and worded completely differently from one for a first-time user. Audience extraction also drives tone, vocabulary ceiling, and assumed knowledge. Success criteria — Extracted for complex tasks where “correct” is not self-evident. Wherever possible, success criteria are converted to binary checks: Does the output contain a concrete example for every claim? Yes/No. Binary criteria produce prompts that are measurably verifiable rather than subjectively good-ish. Examples — Extracted and embedded as few-shot pairs when output format is critical and ambiguous. If you describe a desired output pattern and it can be expressed as a concrete before/after pair, that pair will appear in the prompt. Pattern lock via examples is significantly more reliable than describing a format in prose.What Happens When Critical Dimensions Are Missing
If the extraction step finds that one or more always-extracted dimensions cannot be resolved, Prompt Master asks. The questions are targeted, specific, and limited to three maximum across the entire interaction.The three-question ceiling is a hard limit, not a guideline. Longer interrogations shift the cognitive burden back to you. If three questions are not enough to resolve the ambiguity, Prompt Master will make a reasonable assumption for any remaining gaps, state that assumption explicitly, and deliver the prompt. You can then correct the assumption in the next turn.
The Token Efficiency Audit
After the intent dimensions are extracted and the prompt is drafted, a six-point audit fires before delivery. This audit is what separates a prompt that looks correct from one that actually works on the first attempt.Is the prompt formatted for the confirmed target tool’s specific syntax and instruction-following patterns? A Claude prompt and a GPT-4o prompt for the same task should look structurally different.
Are the most critical constraints positioned in the first 30% of the prompt? Constraints buried at the end of a long prompt are frequently ignored due to attention decay in transformer architectures.
Does every instruction use the strongest available signal word?
MUST instead of should. NEVER instead of avoid. ALWAYS instead of try to. Weak signal words give the model permission to deviate.Have all excluded techniques been removed? Mixture of Experts, Tree of Thought, Graph of Thought, Universal Self-Consistency, and layered prompt chaining are permanently excluded because they increase fabrication risk.
Is every sentence doing work? Vague adjectives are cut. Format instructions are made explicit. Scope is bounded by hard delimiters rather than polite suggestions.
Example: Dimension Extraction in Practice
Here is how the extraction step processes a real request. Raw request: “Can you help me write a prompt for summarizing customer feedback?”| Dimension | Extracted value | Notes |
|---|---|---|
| Task | Summarize customer feedback entries into structured insight reports | ”Help me write” → the actual task is summarization; “customer feedback” → input domain confirmed |
| Target tool | ❓ Not specified | Will trigger one clarifying question |
| Output format | ❓ Not specified | Length, structure, and filetype unknown — will trigger one clarifying question |
| Constraints | None detected | No prohibitions mentioned |
| Input | Customer feedback text (format unknown) | Implied by “customer feedback” |
| Context | None | No session history |
| Audience | Not specified | Will default to internal/analytical unless user-facing context emerges |
| Success criteria | Not specified | Will default to: output captures sentiment, key themes, and actionable items |
| Examples | None provided | No input/output pairs given |
- Which AI tool will this prompt be used in? (e.g., Claude, ChatGPT, Gemini)
- What output format do you need — for example, a bullet-point list per entry, a structured JSON object, or a prose paragraph summary?