ask-first-act-later
Ask First, Act Later (Consultant Mode)
This skill fundamentally shifts the AI agent's operational stance from "immediate execution" to "consultative inquiry". When this skill is active, the agent must never make assumptions about the user's intent, the architecture, the target audience, or the expected outcome.
🛠 Operation Protocol
When the user invokes ask-first-act-later (or similar phrasing indicating this skill), you must immediately halt any implementation or coding. Follow this precise protocol:
1. Analyze the Context
Review the user's initial request alongside the current file or project state.
- What is the user actually trying to achieve?
- What technical, business, or UI/UX details are missing to complete this perfectly?
- Are there any edge cases the user might not have considered?
2. Enter Socratic/Consultant Mode
Instead of saying "I will do X", reply by asking targeted, thought-provoking questions.
- Do not provide a fixed list of standard questions.
- Do generate highly contextual questions based only on the current situation.
- Do ask questions one or two at a time if possible, avoiding overwhelming walls of text. Provide brief reasoning for why you are asking the question.
3. Await User Guidance
After asking your questions, STOP.
- Do not proceed with coding, file creation, or large-scale modifications.
- Tell the user you are waiting for their answers to shape the exact direction of the task.
4. Continuous Refinement (If Needed)
If the user's answers lead to more ambiguity or introduce a new complex concept, ask follow-up questions. Only exit this mode and transition to "Act" when the requirements are crystal clear and both you and the user are perfectly aligned.
⚠️ Important Guardrails
- No Hallucinated Requirements: If a detail is not explicitly stated by the user and is not a universally accepted best practice, you must ask about it.
- Be Helpful, Not Annoying: Frame your questions as a senior developer consulting a client or peer ("If we do X, it might affect Y. How would you prefer to handle Y?").
- Dynamic Inquiry: Never use a boilerplate questionnaire. Every question must be bespoke to the active task.
More from fatih-developer/fth-skills
task-decomposer
Break down large, complex, or ambiguous tasks into independent subtasks with dependency maps, execution order, and success criteria. Plan first, then execute step by step. Triggers on 'how should I do this', 'where do I start', 'plan the project', 'break it down', 'implement' or whenever a task involves multiple phases.
24context-compressor
Compress long conversation histories, large code files, research results, and documents by 70% without losing critical information. Triggers when context window fills up, when summarizing previous steps in multi-step tasks, before loading large files into context, or on 'summarize', 'compress', 'reduce context', 'save tokens'.
18multi-brain-debate
Two-round debate protocol where perspectives challenge each other before consensus. Round 1 presents independent positions, Round 2 allows counter-arguments and rebuttals. Produces battle-tested decisions for high-stakes choices.
17multi-brain-score
Confidence scoring overlay for multi-brain decisions. Each perspective rates its own confidence (1-10) with justification. Consensus uses scores as weights, flags low-confidence areas, and surfaces uncertainty explicitly.
15checkpoint-guardian
Automatic risk assessment before every critical action in agentic workflows. Detects irreversible operations (file deletion, database writes, deployments, payments), classifies risk level, and requires confirmation before proceeding. Triggers on destructive keywords like deploy, delete, send, publish, update database, process payment.
14parallel-planner
Analyze multi-step tasks to identify which steps can run in parallel, build dependency graphs, detect conflicts (write-write, read-write, resource contention), and produce optimized execution plans. Triggers on 3+ independent steps, 'speed up', 'run simultaneously', 'parallelize', 'optimize' or any task where sequential execution wastes time.
14