Agent-Skills.md

Agent Skills: Socratic Plan Interview

Socratic interview that stress-tests a plan against project docs and code, one question at a time. Use when user says "grill my plan", "challenge this", "stress-test", or wants to reach shared understanding before implementing.

UncategorizedID: tianjianjiang/smith/smith-dialectic

Repository

tianjianjiang/smith
tianjianjiang
1

Install this agent skill to your local

pnpm dlx add-skill https://github.com/tianjianjiang/smith/tree/HEAD/smith-dialectic

Skill Files

Browse the full folder contents for smith-dialectic.

Download Skill

Loading file tree…

smith-dialectic/SKILL.md

Skill Metadata

Name
smith-dialectic
Description
Socratic interview that stress-tests a plan

Socratic Plan Interview

Scope: Relentless Socratic questioning of user's plan against project docs and code Load if: User requests plan challenge, stress-test, dialectic, or "grill" Prerequisites: @smith-guidance/SKILL.md (Questioning Techniques, Anti-Sycophancy) Based on: mattpocock/skills grill-with-docs

CRITICAL: Interview Protocol

The interview targets the USER's plan, never a third party's work. To red-team FINDINGS instead — your own or a third party's — use @smith-validation Adversarial Verification, not this interview.

Interview the user relentlessly about every aspect of their plan until reaching shared understanding. Walk down each branch of the design decision tree, resolving dependencies one by one.

For each question:

  1. Ask ONE question at a time
  2. Provide your recommended answer
  3. Wait for user feedback before continuing
  4. If the question can be answered by exploring the codebase, explore the codebase instead of asking

Before starting, confirm scope:

  • What plan/design is being stress-tested?
  • Which project docs to use as evidence? (default: CLAUDE.md + README)

Keep the interview scoped and adversarially honest:

  • Evaluate only the user's own plan, never a third party's work
  • Ask open-ended Socratic questions rather than using checklist-audit pass/fail grading or scoring rubrics
  • Confirm the user's objective before expanding scope
  • Challenge substantively rather than rubber-stamping ("looks good")
  • Ask one question per response, not a batch

During the Interview

Challenge against project docs: When the user's plan conflicts with CLAUDE.md, README, or project conventions, surface it immediately. "CLAUDE.md says X, but your plan assumes Y — which is right?"

Sharpen fuzzy language: When the user uses vague or overloaded terms, propose a precise canonical term. "You're saying 'service' — do you mean the API layer or the background worker? Those have different constraints."

Discuss concrete scenarios: When relationships or dependencies are discussed, stress-test with specific scenarios. Invent edge cases that force the user to be precise about boundaries.

Cross-reference with code: When the user states how something works, check whether the code agrees. Surface contradictions: "Your code does A, but you just said B — which is current?"

Questioning Style

Builds on @smith-guidance/SKILL.md Socratic Method:

  • Clarify: "What exactly do you mean when you say X covers Y?"
  • Challenge: "The README says this requires A. Your plan assumes B. How do you reconcile?"
  • Evidence: "Which doc supports this? I checked Z and found no mention."
  • Implications: "If this assumption is wrong, what breaks downstream?"

Maintain adversarial posture. Apply smith-guidance Anti-Sycophancy rules throughout.

Related

  • @smith-guidance/SKILL.md - Socratic Method, Anti-Sycophancy (foundation)
  • @smith-analysis/SKILL.md - Constructive reasoning (complement)
  • @smith-clarity/SKILL.md - Cognitive traps (defensive complement)
  • @smith-validation/SKILL.md - Hypothesis testing; adversarial verification of findings

Before You Finish

When conducting a dialectic session:

  1. Confirm user's plan and doc scope before starting
  2. Read all relevant project docs + explore code
  3. Walk the design decision tree branch by branch
  4. One question at a time, with your recommended answer
  5. Challenge language, assumptions, and claims against evidence
  6. Continue until shared understanding is reached