Target
$ARGUMENTS
Initial Alignment Check
Verify your understanding matches the user's intent through detailed questions.
Purpose
LLMs can drift from user intent during implementation. This catches misalignment early by asking specific questions about what was built.
Question Density
- •Per 2-3 functions: 1-2 questions
- •Per file: 1-2 questions about the file's role
- •Overall: Questions about how pieces connect
Question Types
Use diagrams and visuals to test understanding - they reveal misalignment faster than words.
Function-Level
- •"This function does X - is that what you intended?"
- •"I made Y return Z in edge case W - correct?"
- •"Function A calls B here - is that the right dependency direction?"
File-Level
- •"This file handles X responsibility - does that match your mental model?"
- •"I put Y in this module because Z - does that make sense?"
Connection-Level (use visuals)
Show the flow and ask if it's right:
code
"Is this the correct data flow?
[User Input] → [Validator] → [Processor] → [Store]
↓
[Error Handler]
"
code
"Does this dependency structure match your intent?
┌──────────┐ ┌──────────┐
│ Service │────>│ Repo │
└────┬─────┘ └──────────┘
│
v
┌──────────┐
│ Cache │
└──────────┘
"
Path-Level (use visuals)
code
"Is this the correct execution path for case X? 1. Entry: handleRequest() 2. Validate: checkAuth() → checkParams() 3. Process: transform() → save() 4. Return: formatResponse() "
Process
This is a cycle, not a one-shot check. Keep going until alignment feels solid.
code
┌─────────────────────────────────────┐ │ │ │ ┌─────────┐ ┌─────────────┐ │ │ │ Ask │───>│ Listen to │ │ │ │Questions│ │ Answers │ │ │ └─────────┘ └──────┬──────┘ │ │ ^ │ │ │ │ v │ │ │ ┌─────────────┐ │ │ │ │ Aligned? │ │ │ │ └──────┬──────┘ │ │ │ │ │ │ │ No │ Yes │ │ └────────────────┘ │ │ │ v │ │ ┌────────┐ │ │ │ Done │ │ │ └────────┘ │ └─────────────────────────────────────┘
- •Review changed code
- •Ask questions about functions, files, connections
- •Listen to answers - do they confirm or correct your understanding?
- •If corrected, ask follow-up questions about the correction
- •Repeat until answers feel aligned - no more surprises
- •Document corrections in deviation file
When to Stop
Stop when:
- •Answers confirm your understanding without corrections
- •Follow-up questions get "yes, exactly" responses
- •You can predict what the user will say
Keep going when:
- •Answers reveal you misunderstood something
- •User adds context you didn't have
- •You're unsure about any part
Output Format
Ask questions using AskUserQuestion tool. Group related questions.
After alignment:
code
## Alignment Summary
- Cycles: {N rounds of questions}
- Corrections needed: {list or "none"}
- Understanding confirmed for: {list of files/areas}
Success Criteria
Full alignment with no surprises is the goal. Keep cycling until you get there.