Prompt Profiles for LLMs: Claude, GPT, Gemini, and More

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.

​

Claude 4.x (Opus 4.8 default / 4.7)

• Front-load everything — put the most important instructions at the top of the prompt, not buried in context
• Use XML tags for complex prompts — wrap sections in <instructions>, <context>, <examples>, <output> for reliable parsing
• Adaptive thinking is built-in — never add explicit chain-of-thought instructions (CoT); Claude 4.x handles reasoning internally
• 1M context window — safe for large codebases and long documents; no truncation anxiety
• Literal instruction following — specify precisely what you want; Claude will not infer unstated preferences
• Append for agentic tasks: Only make changes directly requested. — prevents Claude from over-engineering solutions
• Template: Template M (agentic tasks), CO-STAR (structured requests), RTF (quick tasks)

<instructions>
Refactor the function below to eliminate duplicate logic.
Only make changes directly requested.
Do not rename variables, add comments, or reformat code outside the changed lines.
</instructions>

<context>
Language: TypeScript
File: src/utils/parser.ts
</context>

<task>
[paste function here]
</task>

​

ChatGPT / GPT-5.x

• Smallest prompt that works — strip every word that doesn’t carry instruction weight
• Explicit output contract — state the exact format, length, and structure you expect up front
• Compact structured outputs — prefer JSON, numbered lists, and tables over free prose responses
• Constrain verbosity with word counts — Respond in under 150 words or Return exactly 3 bullet points prevents runaway outputs
• Strong at long-context synthesis — safe to provide large documents for summarization, analysis, or extraction tasks

You are a code reviewer. Return a JSON array of issues.
Each issue: { "line": number, "severity": "error"|"warning"|"info", "message": string }
Respond with the array only — no prose, no markdown fences.
Max 10 issues. Prioritize by severity descending.

​

o3 / o4-mini

• SHORT, clean instructions only — keep the full prompt concise; clarity beats completeness here
• System prompts under 200 words — longer system prompts reduce reasoning quality
• Zero-shot first — do not add few-shot examples unless the task genuinely requires them; examples constrain reasoning
• No CoT scaffolding — never add step-by-step reasoning instructions (see warning above)
• State the goal, not the method — describe what you want, not how to get there

Review the attached code for security vulnerabilities.
Return: vulnerability name, affected line numbers, severity (critical/high/medium/low), fix recommendation.
Format as a markdown table.

​

Gemini 2.x / 3

• Strong long-context and multimodal capability — safe for document-heavy, image-paired, and mixed-media tasks
• Add a grounding anchor — explicitly instruct Only cite sources provided in this prompt or Do not fabricate citations to prevent hallucinated references
• Explicit format locks — state the output structure precisely; Gemini drifts into verbose prose without a hard format constraint
• Avoid open-ended “discuss” prompts — prefer specific deliverables (return a table, write exactly 5 points)

Analyze the research paper below and extract key findings.
Format: numbered list, max 8 items, one sentence each.
Only reference claims that appear explicitly in the provided text.
Do not add citations, context, or background not present in the document.

[document text]

​

Qwen 2.5

• Clear, explicit system prompt — Qwen 2.5 uses the system prompt effectively; invest in it
• Shorter, focused prompts — concentrated task descriptions outperform sprawling multi-part instructions
• Strong at JSON and structured output — the safest choice for structured extraction tasks when using open-weight models
• Explicit instruction following — state each requirement as a separate line item rather than embedding in prose

System: You are a data extraction assistant. Return only valid JSON. No prose.

Extract the following fields from the customer email below:
- sender_name
- issue_category (billing | technical | general)
- urgency (high | medium | low)
- one_line_summary

Email: [paste email]

​

Qwen3 (Thinking vs. Non-Thinking Mode)

Identify the most likely root cause of the bug described below.
Return: root cause (one sentence), affected component, suggested fix.

Bug report: [paste report]

System: You are a JSON-only assistant. Return valid JSON, no other text.

Classify the support ticket below:
- category: billing | technical | account | other
- priority: 1 (critical) to 5 (low)
- escalate: true | false

Ticket: [paste ticket]

​

Ollama

• Always identify the underlying model first — mistral, llama3, gemma2, qwen2.5, deepseek-r1 all have different optimal strategies
• System prompt is the most impactful lever — Ollama models respond strongly to a well-crafted system prompt; invest here first
• Temperature 0.1 for coding tasks — deterministic output for code generation, debugging, and data extraction
• Temperature 0.7–0.8 for creative tasks — writing, brainstorming, ideation, dialogue

System: You are a senior Python developer. Write clean, PEP-8-compliant code.
Return only the function — no explanation, no markdown fences.

Task: Write a function that validates an email address using regex.
Signature: def validate_email(email: str) -> bool

​

Llama / Mistral

• Shorter, simpler prompts — concise beats comprehensive; long prompts degrade instruction-following
• Simple flat structure — avoid nested XML, elaborate JSON schemas, or multi-section templates
• More explicit than Claude or GPT — don’t rely on inference; state each requirement directly
• Avoid implicit assumptions — what Claude infers from context, Llama needs stated explicitly

Write a Python function that sorts a list of dictionaries by the "date" key.
- Input: list of dicts, each with a "date" key in "YYYY-MM-DD" format
- Output: sorted list, ascending order
- Handle missing "date" keys by placing those items at the end
- Return only the function, no explanation

​

DeepSeek-R1

• Reasoning-native — no CoT instructions — same rule as o3; short, clean goal statements only
• Short, clean instructions — let the model’s internal reasoning do the work
• Outputs appear in <think> tags — final answer follows the closing tag; parse accordingly if using the API
• Strong at math, logic, and complex reasoning — well-suited for multi-step analytical tasks

Determine whether the following business logic contains a race condition.
If yes: identify the affected operations and propose a fix.
If no: confirm with a one-sentence explanation.

Code: [paste code]

​

MiniMax M3 / M2.7

• OpenAI-compatible API — works with the same message format as GPT; system/user/assistant roles apply
• 1M context window on M2.7 — safe for very large document analysis and codebase tasks
• Temperature must be between 0 and 1 — values outside this range will error; do not use 0 as a floor if the API requires a positive value
• Strong at code, JSON, and multi-step workflows — reliable structured output and sequential task handling
• Treat prompting style like GPT-5.x — explicit output contract, constrained verbosity, compact structure

System: You are a JSON extraction assistant. Return only valid JSON. No markdown, no explanation.

From the contract excerpt below, extract:
{
  "party_a": string,
  "party_b": string,
  "effective_date": "YYYY-MM-DD",
  "termination_clause": boolean,
  "governing_law": string
}

Contract: [paste excerpt]

​

At a Glance: CoT Rules by Model