unformal/Docs

Best Practices

Practical playbooks for getting the most out of Unformal. The difference between a Pulse that collects noise and one that produces signal is mostly in how you set it up.

Writing a great intention

The intention is the most important field in a Pulse. It's what the AI reads to understand what conversation to have. A vague intention produces a vague conversation. A precise one produces sharp, useful output.

The formula

1

What to learn

The specific information you want to extract

2

Who will answer

Give the AI context about the respondent

3

What it informs

The decision or action this data will feed into

Avoid

"Collect information from potential clients"

Good

"Understand what a new architecture client needs before our first meeting — their budget, project scope, timeline, and what a successful outcome looks like for them. Use this to prepare a tailored proposal."

Avoid

"Get feedback on our product"

Good

"Gather structured feedback from beta users about their experience with the onboarding flow. I want to know where they got confused, what they loved, and what's stopping them from using the feature daily. This feeds into our next sprint planning."

  • Be specific about what you want to learn — the more concrete, the better the questions
  • Include who will be answering ("potential client", "beta user", "internal team member")
  • Mention the downstream decision: what will you do with this information?
  • Aim for 2–4 sentences. Too short = vague. Too long = the AI gets confused about priorities.
💡

Use the ✨ Generate button in the Pulse creator. Type your intention, click Generate, and the AI will suggest a tone, topics, and output fields. It's the fastest way to get a good starting point.

Defining effective output fields

Output fields are how you tell the AI what structured data to extract from each conversation. They directly determine the quality and consistency of your Echo data.

Choose the right type

string

Qualitative data, descriptions, free-text answers

"What's their biggest pain point?"

number

Quantitative values, ratings, counts, durations

"Timeline in weeks", "Budget in EUR"

boolean

Yes/no questions, flags, presence/absence

"Are they the decision maker?"

array

Lists of items, multiple options, ranked preferences

"Top 3 challenges", "Tools currently used"

Write descriptions the AI can act on

The description isn't just for you — the AI reads it to know what to extract. Vague descriptions produce vague extractions.

Vague field

"budget"

Clear field

"budget_range — Estimated annual budget range the respondent is working with, as a string like '€10k–50k' or 'under €5k'"

Too generic

"notes"

Specific and actionable

"decision_blockers — Array of specific concerns, objections, or blockers preventing the respondent from moving forward right now"

Sweet spot: 8–12 fields

Too few fields and you miss important data. Too many and the AI tries to cram every topic into a short conversation, producing shallow answers.

< 5 fields

Likely missing important dimensions

8–12 fields

The sweet spot for most use cases

> 15 fields

Conversations feel rushed and shallow

  • Use snake_case for names: budget_range not Budget Range
  • Mix types: some strings for qualitative, some numbers for quantitative, arrays for lists
  • Write descriptions in plain English — imagine explaining to a junior analyst what to look for
  • Avoid overlapping fields — if two fields could capture the same thing, merge them

🎙️ Interview mode tips

Interview mode works best when you want depth. The AI asks one question at a time, listens, and decides what to ask next. Here's how to configure it well.

Setting maxQuestions

This controls how long the conversation runs. The AI tries to use all available questions to cover your topics and follow interesting threads.

3–5 questionsQuick pulse checks, NPS follow-up, short feedback
8–12 questionsClient intake, user research, most use cases
15–20 questionsDeep discovery, research interviews, complex intake
💡

Keep maxQuestions between 8–12 for most use cases. This gives the AI room to follow interesting threads without making respondents feel like they're being interrogated.

Defining topics

Topics tell the AI what areas to cover. Without topics, the AI makes reasonable guesses from your intention — but topics give you control over coverage.

// Good topics for a client intake Pulse

"topics": [
  "Current situation and main challenges",
  "Budget and decision-making process",
  "Timeline and urgency",
  "What success looks like",
  "Previous solutions tried",
  "Team size and context"
]
  • Define 5–7 topics for an 8–12 question conversation
  • Write topics as themes, not specific questions — let the AI phrase them naturally
  • Order topics roughly: context first, specifics later, goals last
  • The AI will follow interesting threads — topics ensure nothing important gets skipped

Tone selection

Match the tone to your respondent, not to your brand.

Talking to enterprise buyers

Professional or Conversational

Creative professionals, architects

Conversational or Casual

Internal team members

Casual or Coaching

Users giving feedback

Conversational or Coaching

🗒️ Extract mode tips

Extract mode flips the dynamic: the respondent writes everything they want to say first, then the AI asks only about what's missing. It's faster, less friction, and works best when respondents already know what they want to share.

When to use Extract mode

Brain dumps

"Just tell me everything" — the AI structures it after

Technical stakeholders

Engineers and PMs prefer writing context themselves, not being led by questions

Bug reports

Users describe the issue, AI extracts: severity, steps to reproduce, expected vs actual behavior

Quick intake

Situations where asking questions one by one feels patronizing

💡

Extract mode is great for users who say "I don't want to answer questions, I just want to write what I know." The AI handles the structuring work so you don't have to pre-structure the intake.

Optimizing Extract mode output fields

In Extract mode, the AI tracks field coverage as the respondent writes. It only asks follow-up questions about fields that weren't addressed. Make sure your output fields cover everything you actually need.

  • Use more fields than in Interview mode — the AI extracts passively from the dump, only asks for gaps
  • Write field descriptions that work even without a direct question being asked
  • Casual or Conversational tone works best — extract mode is inherently informal
  • The welcome screen text matters: tell respondents "write everything you know" — the more they write, the fewer follow-ups needed

Use case examples

Real configurations for common scenarios.

🤝

Client intake

InterviewProfessional10 questions

Intention

"Understand what a new client needs before our first meeting — budget range, project scope, timeline, and what success looks like. Use this to prepare a proposal."

Key output fields

budget_range · project_type · timeline · decision_maker · current_solution · pain_points (array) · success_criteria · next_steps

🔬

User research

InterviewConversational12 questions

Intention

"Learn how beta users experience the onboarding flow. Find where they get confused, what clicks, and what stops them using the feature daily. Feeds into next sprint."

Key output fields

first_impression · confusion_points (array) · aha_moment · daily_blocker · feature_rating (number) · missing_features · recommendation_likelihood (number)

🧠

Brain dump intake

ExtractCasual8 questions max

Intention

"Get all the context from a colleague about a project they've been running — status, blockers, key decisions made, what still needs to happen. I'm taking it over and need to get up to speed quickly."

Key output fields

project_status · key_decisions (array) · active_blockers (array) · next_steps (array) · stakeholders (array) · context_notes

Welcome screen copy: "Write down everything you know about this project — don't worry about structure. The AI will organize it and ask about anything missing."