plan-interview

Plan Interview Skill

Install

gh skill install pskoett/pskoett-skills plan-interview

Fallback using the Agent Skills CLI:

npx skills add pskoett/pskoett-skills/skills/plan-interview

Philosophy

Every skill in this collection is built around a core philosophy — a principle that agents struggle to internalize on their own.

This skill's philosophy: "Make the change easy, then make the change."

Agents default to plowing straight through implementation, no matter how tangled the path. They rarely pause to ask: "Would a preparatory refactor make this change simple instead of hard?" Planning is where that question gets asked. During codebase exploration and plan generation, actively look for structural friction — code that makes the target change awkward, brittle, or overly complex. When you find it, the plan should propose a preparatory step: make the change easy first, then make the change itself. Two clean steps beat one heroic slog.

Purpose

Run a structured requirements interview before planning implementation. This ensures alignment between you and the user by gathering explicit requirements rather than making assumptions.

When Invoked

User calls /plan-interview <task description>.

Skip this skill if the task is purely research/exploration (not implementation).

Interview Process

Phase 1: Upfront Interview (Before Exploration)

Check for the host's structured question tool (ask_user, AskUserQuestion, or an equivalent). Use it for focused questions that the user can answer before the next planning step. In hosts without a structured question tool, ask the same question directly in chat and pause for the response. Do not fall back to plain chat merely because the host uses a different tool name. Respect the host schema: if the tool accepts only one question per call, ask the domains sequentially rather than bundling them.

Required Question Domains

Cover ALL four domains before proceeding:

1. Technical Constraints

   • Performance requirements
   • Compatibility needs
   • Existing patterns to follow
   • Architecture understanding (if codebase is unfamiliar)

2. Scope Boundaries

   • What's explicitly OUT of scope
   • MVP vs full vision
   • Dependencies on other work

3. Risk Tolerance

   • Acceptable tradeoffs (speed vs quality)
   • Tech debt tolerance
   • Breaking change acceptance

4. Success Criteria

   • How will we know it's done?
   • What defines "working correctly"?
   • Testing/validation requirements

Question Generation

• Generate questions dynamically based on the task - no fixed template
• Group related questions into thematic batches
• 2-3 questions per batch (do not exceed)
• Continue until you have actionable specificity (can describe concrete implementation steps)

Planning Depth Calibration

Before leaving the interview phase, classify the task and choose a planning depth:

• Simple/trivial (small bug fix, isolated change): minimal plan, at most 1 refinement pass
• Moderate (feature work in known area): standard plan, usually 1-2 refinement passes
• Complex/high-risk (multi-file, new architecture, unfamiliar codebase, migrations, auth, concurrency): deep plan with iterative refinement until improvements flatten

Let the user override this (fast vs deep) if they have a clear preference.

Handling Edge Cases

Anti-Patterns to Avoid

• Do NOT ask variations of the same question
• Do NOT make major assumptions without asking
• Do NOT over-engineer plans for simple tasks

Phase 2: Codebase Exploration

After interview completes, explore the codebase to understand:

• Existing patterns relevant to the task
• Files that will be affected
• Integration points
• Potential risks
• Structural friction — Is the current code shape fighting the planned change? Would a preparatory refactor (rename, extract, restructure) make the actual implementation straightforward? If yes, propose it as a distinct first step in the plan.

For complex or unfamiliar projects, do a brief context refresh before deep planning:

• Re-read AGENTS.md and README.md if present and relevant
• Identify the current architecture boundaries and conventions before refining the plan
• If the session was interrupted or context drifted, refresh these again before another refinement round

Knowledge Audit (Between Exploration and Planning)

Before writing the plan, explicitly ask: "Does the knowledge needed to complete this task exist somewhere I can reach?"

For each significant implementation step, classify where the required knowledge lives:

• Codebase — Existing patterns, conventions, or code that demonstrates how to do it. Found during exploration.
• Prompt/context — User-provided requirements, constraints, or domain knowledge from the interview.
• Training data — General programming knowledge, well-known libraries, standard patterns the model reliably knows.
• Nowhere reachable — The knowledge isn't in any of the above. The agent would be guessing.

When a step falls into "nowhere reachable":

1. Stop and surface it. Do not fill the gap with confident-sounding guesses.
2. Ask the user to provide the missing knowledge, point to documentation, or confirm that best-effort is acceptable.
3. If the user provides a reference (docs URL, API spec, example), load it before planning that step.
4. If the gap cannot be filled, mark the step as blocked in the plan's Open Questions section.

This audit prevents the most common agent failure: confidently proceeding when the knowledge simply isn't there, producing plausible but wrong output.

Phase 3: Plan Generation

Write plan to docs/plans/plan-NNN-<slug>.md where NNN is sequential.

Use a draft -> refine workflow. Stay in plan space while you are still finding material improvements. Planning tokens are usually much cheaper than implementation tokens for non-trivial work.

Draft First, Then Refine

1. Create a draft plan from the interview + exploration results.
2. Run iterative refinement passes before asking for approval (depth based on task complexity).
3. Present the refined plan for user review.

Iterative Plan Refinement Loop (Before User Review)

Run 1..N refinement passes depending on complexity. For each pass:

1. Fresh-eyes start (mandatory): Re-read the interview answers, constraints, success criteria, and the current draft plan with "fresh eyes" before revising anything.
2. Check for contradictions, missing edge cases, integration risks, and vague implementation steps.
3. Improve architecture, sequencing, and reliability where it clearly helps users.
4. Strengthen the testing and validation plan (unit + integration/e2e where applicable, plus useful diagnostics/logging).
5. Verify feature preservation:
   • Do NOT oversimplify
   • Do NOT remove agreed features or functionality unless the user explicitly approves a scope reduction
6. Record a short per-pass summary: what changed and why.

Stop iterating when any of the following is true:

• Two consecutive passes produce no material improvements
• Changes are only wording/style with no effect on execution quality
• The task is simple and the plan is already actionable
• The user asks to stop and proceed

Optional: Multi-Plan Synthesis ("Best of All Worlds")

If the user provides multiple competing plans (from different models or prior iterations):

• Compare them honestly against the current plan
• Extract the best ideas, tradeoffs, and risk mitigations
• Merge them into a single canonical plan that preserves agreed scope
• Prefer showing git-diff style changes to the existing plan when the user asks for revision output

Reusable prompt templates for the refinement loop and multi-plan synthesis live in references/iterative-plan-refinement-prompts.md.

Required Elements

Every plan MUST include:

Optional Elements

Include when relevant:

• Rejected Alternatives: Only for major architectural decisions
• Decision Tree: Only when multiple valid approaches exist
• Visual Diagrams: ASCII or Mermaid when helpful for understanding

Constraints

• No time estimates - describe what needs doing, not how long
• No length limits - plan should match task complexity
• No silent scope reduction - do not drop agreed features to make the plan "cleaner"
• Don't over-iterate simple work - use the planning depth calibration above
• Freeform structure beyond required elements

Phase 4: Post-Approval

When user approves the plan:

1. Auto-start implementation immediately (no separate "proceed" confirmation needed). If intent-framed-agent is active, flow directly into a faithful Intent Frame and reuse the plan approval. Ask again only if the frame introduces a material change or unresolved decision.
2. Populate TodoWrite with checklist items (if TodoWrite is not available, track progress via structured comments in your output)
3. At natural breakpoints (significant decisions), compare progress to plan

Fast Mode

If user wants quick planning, use draft + refine:

1. Perform task-focused codebase search
2. Generate draft plan
3. Run abbreviated interview to refine
4. Run exactly one fresh-eyes refinement pass (preserve functionality, tighten steps, add test/validation coverage)

Resume Support

If a partial plan exists in docs/plans/:

If resuming refinement, first summarize the current plan state and the most recent refinement changes, then continue with the fresh-eyes refinement loop.

Interoperability with Other Skills

What this skill consumes

• From the user: Task description, requirements, and answers to interview questions.
• From the codebase: Existing patterns, architecture, and conventions discovered during exploration.
• From context-surfing handoff files (on resume): If a previous session exited due to context drift, the handoff file in .context-surfing/ provides the re-entry point and remaining work for replanning.

What this skill produces

• Plan file (docs/plans/plan-NNN-<slug>.md) — consumed by intent-framed-agent as context for the intent frame, and by context-surfing as part of the wave anchor. Copied verbatim into handoff files on drift exit.
• Planning depth classification — informs how many skills to activate (see pipeline depth table in README).

Pipeline position

1. plan-interview (requirements and plan generation — you are here)
2. intent-framed-agent (execution contract + scope drift monitoring)
3. context-surfing (optional context-quality monitoring for Large, Long-running, or explicitly context-sensitive execution)
4. verify-gate (machine verification: compile + test + lint)
5. simplify-and-harden (post-completion quality/security pass)
6. self-improvement (capture recurring patterns and promote durable rules)

Example