Step1

/step1 - Explore → Interrogate → Bridge

Structured Q&A that converges on decisions. Questions evolve based on answers.

Arguments

• /step1 — asks "What needs solving?"
• /step1 <topic> — skips opening question
• /step1 resume — finds last decision tracker in conversation, continues

Flow

1. EXPLORE

1. Read relevant codebase files (5-10). For greenfield projects, ask about constraints instead.
2. Identify decision points. Form an opinion on each from codebase evidence.
3. Output: "~N decisions to work through (some will be auto-decided)"

2. INTERROGATE

One question at a time via AskUserQuestion.

Auto-decide when:

• Codebase evidence makes one option overwhelmingly correct
• LLM-readability clearly favors one option (naming, structure, indirection) — still ask, but put the best option first with (Recommended) and include the rationale so the user can confirm quickly

Present initial auto-decisions as a batch before the first human question:

### Auto-decided (review before we continue)
1. ✅ (auto) [topic]: [answer] — [rationale]

Ask: "Want to revisit any of these, or move on?" If flagged, convert to regular question.

Questions:

• Order options by recommendation strength (best first)
• All option slots are for real answers — user can type "go back" or "park" via Other

After each answer:

• Reassess: does this reveal new decisions (mark 🆕), kill existing ones (mark ➖)?
• Flag conflicts with previous decisions
• Push back on risky choices even if user seems confident
• If decisions exceed 8, suggest splitting the task before continuing
• Output tracker:

### Decisions (N decided, ~M remaining)
1. ✅ [topic]: [answer] — [why]
2. ✅ (auto) [topic]: [answer] — [rationale]
3. ⬜ [next]...

Go back (user types via Other): Re-ask previous question, mark it ⬜, re-evaluate downstream.

Park (user types via Other): Stop. Output tracker as-is. Write _step1_decisions.md with ## Unresolved section. Do NOT suggest /step2.

3. BRIDGE

Early resolution — ONLY when outcome is a single localized fix (one file, no conventions, no docs):

• Skip _step1_decisions.md, summarize inline. Simple enough to implement directly — no pipeline needed.
• After the summary, ask: "Ready to implement? I can do this directly." Do NOT suggest /step2 — early resolution bypasses the pipeline entirely.
• Question count is irrelevant. 1 human question + multi-file scope = full bridge.

Full bridge — default path (multi-file, conventions, or docs):

1. Present decision block(s):

## Decision: [topic]
**Approach:** [what and how]
**Why:** [codebase-grounded reasons]
**Rejected:** [alternative] — [why it fails]; ...
**Risks:** [trade-offs and pitfalls carried by chosen approach]
**Scope:** [in/out]
**Key decisions:**
- [decision]

Rejected and Risks exist so /step2 can convert them into guardrails. One line each. Omit if genuinely empty. 2. Ask "Anything you'd change?" — apply changes if requested. 3. MANDATORY: Write _step1_decisions.md at repo root using the Write tool. Do this BEFORE any closing remarks. This file is the primary deliverable — without it, /step1 failed. 4. Only AFTER the file is confirmed written, end with: Next: /step2 to plan the implementation. Tip: /clear first so /step2 gets a full context window.

Rules

• NEVER use Task or EnterPlanMode. This skill IS the exploration framework. Subagents lose your context. Plan mode hijacks your flow. Read/Glob/Grep directly.
• No implementation. Only output is _step1_decisions.md. Exception: early resolution.
• One question at a time. Never batch.
• Never auto-chain to /step2.
• Recommend, warn, push back, challenge.
• Take sides. When one option is obviously better for LLM development (explicit names, less indirection, flatter structure), lead with it as (Recommended) with rationale. Still ask — but make the right choice effortless to confirm.
• Gate: file before farewell. Never mention /step2 or signal completion until _step1_decisions.md has been written with the Write tool in the current conversation. The file is proof the skill finished. No file = skill failed.