Interview for Automation

Phase 2: Map the Process Step by Step

Walk the user through the full sequence. For each step:

Core sequence questions:

• "What do you do first?"
• "Then what?" (keep asking until you reach the end)
• "How do you know this step is done?"

Depth probes — use when a step is vague:

• "Can you walk me through that more concretely?"
• "What does that actually look like — what are you clicking, writing, or looking at?"

Decision point probes — use when a step involves judgment:

• "How do you decide between X and Y here?"
• "What signals tell you to go one direction vs. another?"
• "Can you give me an example where you'd choose differently?"

Phase 3: Surface Edge Cases and Quality Standards

After mapping the happy path:

• "What can go wrong at each step? How do you handle it?"
• "Are there situations where you do this differently than normal?"
• "What mistakes do people make when they try to do this?"
• "What tells you the output is good vs. just acceptable?"
• "Are there things you always include or never do?"

Phase 4: Verify and Close Gaps

Before synthesizing:

• "Is there anything we haven't covered that someone would need to know to do this well?"
• "If I tried to follow everything you've told me, where would I get stuck?"

Restate your understanding of the full process in 3–5 sentences and ask the user to correct anything that's off.

Phase 5: Synthesize the Automation Prompt

Once the interview is complete, produce a structured prompt using this template:

# [Task Name] — Automation Prompt

## Role
You are [specific role that matches the expertise and tone needed for this task].

## Context
[1–2 sentences explaining what this task is, why it matters, and what domain it lives in]

## Inputs
- [Input 1]: [what it is and where it comes from]
- [Input 2]: ...

## Process

### Step 1: [Name]
[Specific, actionable instruction. Not "review it" — "check for X, Y, Z."]

### Step 2: [Name]
[Instruction]
- If [condition]: [what to do]
- If [condition]: [what to do]

### Step N: ...

## Quality Standards
- [Standard 1 — measurable or observable, not vague]
- [Standard 2]

## Edge Cases
- If [situation occurs]: [how to handle it]
- If [situation occurs]: [how to handle it]

## Output Format
[Exact description of what the final output should be — format, length, structure, tone]

Interview Principles

Ask one question at a time. Multiple questions get shallow, partial answers.

Follow the user's specificity, not a rigid script. If they go deep on step 3, stay there before moving on.

Prefer concrete over abstract. "I review it for quality" → ask "What specifically are you looking for?"

Capture exact language. If the user says "it needs to feel natural" or "I just know when it's right," include that phrasing in the prompt — it encodes judgment.

Don't assume. Even obvious-sounding steps should be confirmed.

Done Criteria

The output prompt should be complete enough that a capable AI model with no prior context could execute the task and produce output the user would approve — without needing to ask follow-up questions. Ask yourself: "If I handed this prompt to someone who had never met this user, would they get it right the first time?"