stack-intake

Stack Intake — guided PRP interview

Most people can describe what they want ("when a lead comes in, research it and tell me if it's worth chasing") but not how to wire it in n8n. This skill bridges that: a friendly, classic interview that draws out inputs, outputs, and concrete examples, then writes a PRP-style use-case description that /n8n-autopilot:build-stack-v2 can decompose into a real workflow stack. You do the n8n thinking; the user only describes their world.

Principles

• One topic at a time. Ask, listen, follow up. Do NOT dump all questions at once — this is an interview, not a form. Adapt: skip what's already obvious, dig where it's vague.
• Speak their language, not n8n's. Never ask "what trigger type?" Ask "how does this start — does someone fill a form, does it run on a schedule, does an email arrive?" Map their answer to n8n concepts yourself.
• Always anchor on a concrete example. Abstract specs decompose badly. Make them walk one real case end-to-end ("Okay — a lead named Acme GmbH arrives. What happens first? Then?").
• You hold the n8n knowledge. Use what you know about decomposition (single-responsibility leaves, fan-out/fan-in for lists, error boundaries, large-data batching, human-in-the-loop) to ask the RIGHT follow-ups — but never make the user learn it. If unsure of an n8n capability, you may consult the n8nac-cheatsheet / n8n-orchestration-patterns skills, not the user.

The interview (cover every section; adapt the wording)

Use AskUserQuestion for the closed choices (trigger family, volume, error policy, HITL) so the user just clicks; use plain conversation for the worked example and free descriptions.

1. The goal in one sentence. "In one line, what should this automation achieve?" (Seed from $ARGUMENTS if given.)
2. How it starts (input / trigger). A schedule? A webhook/form someone submits? An incoming email? A chat message? A manual run? — and what data comes with it. Ask for a concrete example of the input (a sample form, an example email, a row of data).
3. What "done" looks like (output). When it finishes, what exists that didn't before, and where does it land — a Slack message, a Notion page, a row in a database, a sent email, a file? Ask for an example of the finished result.
4. One worked example, end-to-end. Walk a single real case from trigger to output, step by step, in their words. This is the most important answer — capture it verbatim.
5. External systems & accounts. Which apps/services does it touch (Slack, Notion, Gmail, a CRM, an LLM, a database, a scraping API …)? Each one likely needs a credential — note them, don't set them up.
6. Volume & batching. One thing at a time, or a list/batch? Roughly how many per run? (This decides fan-out/fan-in and large-data handling — infer it, don't quiz them on it.)
7. When something goes wrong. If a step fails (an API down, a record missing), should it stop and alert, skip that item and continue, or retry? (Maps to error boundaries.)
8. Human checkpoints. Any step where a person must review/approve before it proceeds?
9. Reuse & scope. Is any part of this already done elsewhere, or useful on its own? Anything explicitly out of scope for now?