prompter: скилл для создания эффективных запросов

--- name: prompter description: Craft effective prompts for subagent delegation. Use this skill when delegating work to subagents to ensure clear, well-structured prompts that produce high-quality results. --- This skill provides a systematic approach to crafting prompts for subagent delegation, based on LLM prompting best practices. ## 7-Step Prompt Crafting Framework ### Step 1: Define Clear Intent & Output State exactly what you want. Be explicit about: - The task objective - Desired output format - Success criteria ``` BAD: "Improve this code" GOOD: "Refactor this function to reduce complexity. Return the refactored code with a brief explanation of changes made." ``` ### Step 2: Provide Context & Motivation Explain WHY the task matters. Context improves performance: ``` <context> This function handles payment processing. Errors here cause customer billing issues, so correctness is critical. </context> ``` ### Step 3: Structure with XML Tags Use XML tags to organize complex prompts unambiguously: ```xml <task> <objective>What needs to be done</objective> <constraints>Limitations and requirements</constraints> <input>The data/code to work with</input> <output_format>Expected format</output_format> </task> ``` ### Step 4: Add Examples (Few-Shot) Include 3-5 relevant, diverse examples wrapped in tags: ```xml <examples> <example> <input>Sample input</input> <output>Expected output</output> </example> </examples> ``` ### Step 5: Specify Role & Behavior Set the agent's role and default behaviors: ```xml <role>You are a security-focused code reviewer specializing in Python.</role> <behavior> - Default to action over suggestion - Flag security issues immediately - Prefer standard library solutions </behavior> ``` ### Step 6: Define Verification Criteria Tell the agent how to verify success: ```xml <verification> Before completing: 1. Run all tests 2. Check for lint errors 3. Verify the function handles edge cases </verification> ``` ### Step 7: Set Return Requirements Specify exactly what to return: ```xml <return> Provide: 1. The modified files 2. Summary of changes 3. Any remaining issues or next steps </return> ``` ## Prompt Template Use this template when delegating to subagents: ```xml <delegation> <role>[Agent's specialized role]</role> <context> [Background information and why this matters] </context> <task> [Clear description of what needs to be done] </task> <constraints> [Limitations, requirements, things to avoid] </constraints> <input> [Files, data, or context to work with] </input> <output_format> [Expected format and structure of response] </output_format> <verification> [How to verify the task was completed correctly] </verification> <return> [What to return when done] </return> </delegation> ``` ## Best Practices Checklist Before delegating, verify your prompt: - [ ] **Clear intent**: Would a new employee understand what to do? - [ ] **Context provided**: Is the "why" explained? - [ ] **Structured with tags**: Are instructions, context, and input separated? - [ ] **Examples included**: Are there 3-5 relevant examples for complex tasks? - [ ] **Role defined**: Does the agent know its specialization? - [ ] **Success criteria**: Can the agent verify its own work? - [ ] **Return format**: Is it clear what to output? ## Common Patterns ### For Code Tasks ```xml <delegation> <role>Senior software engineer specializing in [language/framework]</role> <task>Implement [specific feature/fix]</task> <constraints> - Follow existing code patterns - No external dependencies without approval - Maintain backward compatibility </constraints> <verification>Run tests and lint</verification> <return>Modified files + summary of changes</return> </delegation> ``` ### For Research Tasks ```xml <delegation> <role>Research analyst</role> <task>Investigate [topic]</task> <output_format> - Summary of findings - Key sources with quotes - Recommendations </output_format> <verification>Cross-reference multiple sources</verification> </delegation> ``` ### For Analysis Tasks ```xml <delegation> <role>Code reviewer</role> <task>Review [files] for [criteria]</task> <output_format> For each issue found: - File and line number - Severity (critical/major/minor) - Description - Suggested fix </output_format> </delegation> ``` ## Anti-Patterns to Avoid 1. **Vague instructions**: "Do something good" → Be specific 2. **Negative-only constraints**: "Don't use X" → Say what TO do 3. **Missing context**: No background → Always explain why 4. **No verification**: No success criteria → Define how to check 5. **Unclear output**: No format specified → Define structure