think-before-coding— agent skill

Overview

AI agents default to writing code immediately — even when the request is ambiguous, the context is incomplete, or multiple valid interpretations exist. This skill forces a deliberate thinking phase before any implementation begins.

The result: fewer rewrites, fewer wrong-direction implementations, and fewer "I assumed X" surprises.

When to Use

Activate this skill before writing any code when:

• The request is ambiguous or underspecified
• Multiple valid implementations exist
• You are unsure about an existing system's constraints
• The task touches security, data integrity, or public APIs
• You feel the urge to "just start coding" to make progress

Process

Phase 1: Understand the Request

1. Read the full request — Do not skim. Re-read once.
2. Identify ambiguities — List every decision you would have to make silently:

• What counts as "success"?
• What are the inputs and expected outputs?
• What should NOT change?
• Are there existing patterns to follow?

1. Surface assumptions — Write them out: "I am assuming X because Y."
2. Check for contradictions — Does the request contradict itself or existing code?

Verify: Can you state the goal in one clear sentence? If not, ask before proceeding.

Phase 2: Clarify Before Acting

1. If genuinely ambiguous, ask a focused clarifying question — not a list of 10 questions. Pick the one blocker.
2. If multiple valid approaches exist, present 2–3 options with tradeoffs, then ask which to proceed with.
3. If something seems wrong with the request, say so directly — don't silently work around it.

Verify: You have a clear, agreed-upon interpretation of the task.

Phase 3: State Your Plan

1. Before writing code, state:

• What you will build
• What you will NOT touch
• What success looks like (your verification criteria)

1. For multi-step tasks, write a brief plan:

   1. [Step] → verify: [check]
   2. [Step] → verify: [check]
   3. [Step] → verify: [check]

Verify: The plan is agreed upon, or you have received approval to proceed.

Phase 4: Code

1. Now write the code — following the plan exactly.

Common Rationalizations (and Rebuttals)

Red Flags

• You're already writing code and you haven't stated what success looks like
• You made a silent assumption about what the user wants
• You chose between two approaches without mentioning the tradeoff
• You're solving a problem the user didn't ask about
• You're 100 lines in and realize you misunderstood the goal

Verification

Before leaving this phase, confirm:

• [ ] The goal is stated in one clear sentence
• [ ] All major assumptions are explicitly named
• [ ] Ambiguities are resolved (or flagged for the user)
• [ ] A brief plan exists with verification criteria per step
• [ ] You know what you will NOT touch

References

• Andrej Karpathy on LLM coding pitfalls
• goal-driven-execution skill
• idea-to-spec skill