Anamap Learning : Core

How to Prompt Anamap AI: Practical Guide

Table of Contents

Why prompts matter (and why vagueness can be okay)

Anamap AI is designed to do deep, exploratory work — including root-cause analysis — when you ask broad, human questions like "Why is revenue down?" That vagueness can be a strength: the agent will generate hypotheses, check possible drivers (traffic, conversion, product mix, attribution changes), and suggest experiments or queries to validate each idea.

That said, the assistant performs best when your Anamap project has meaningful attributes, events, and views filled out. Those data-layer definitions give the agent precise handles it can use when building queries, charts, or recommendations.

So: be creative and ask hard questions. The AI will try to fill in gaps, but providing additional context (attributes, events, date windows) improves precision and lets you get runnable queries and charts right away.

Flexible prompt patterns

You don't need a rigid template to get useful answers. Here are a few flexible patterns you can use depending on how much time you want to spend composing the prompt.

  • Minimal / human prompt (fast): "Why is revenue down?"
  • Guided exploratory prompt (better): "Why is revenue down for the last 30 days? Please analyze traffic, conversion, and product mix, list the top 3 hypotheses, and suggest 3 checks (with queries) to validate each hypothesis. Show a small chart if it helps."
  • Precise/actionable prompt (most specific): "Context: Company Acme, attributes: user_tier (free, pro, enterprise), country, product_category. Task: Compare revenue month-over-month for last 60 days; show a line chart of total revenue and a bar chart breaking revenue down by country for the most recent week. Provide SQL/GA4/Amplitude query examples and a one-paragraph summary of likely causes."

You can mix-and-match these patterns — start with a short question and follow up with clarifying requests.

Short prompts that work

Short human prompts are fine because the assistant will ask clarifying questions or propose checks. Use short prompts when you want an exploratory starting point:

  • "Why are conversions down this week?"
  • "Which user segments had the largest revenue decline last month?"
  • "Forecast revenue for next quarter and explain assumptions."

If you want more precise outputs, attach one-line context after the question (attributes, date range, or the view name).

Examples you can copy — including charts & forecasts

  1. Root-cause analysis (open-ended)

"Why is revenue down? Please provide:

  • Top 5 hypotheses ranked by plausibility
  • For each hypothesis, a short check (query or metric to inspect)
  • One recommended action for each hypothesis Format: numbered list + suggested chart types"

Example assistant outputs might include a time series line chart of revenue, a conversion funnel chart, and a bar chart of revenue by country or product.

  1. Ask for specific chart types

"Task: Show revenue trends for the last 90 days as a line chart (7-day smoothing). Also show revenue by country as a stacked bar chart for the latest week. Explain any notable changes in 3 bullets."

  1. Forecasting request

"Task: Forecast total revenue for next quarter using simple time-series (ARIMA or exponential smoothing). Provide the forecast numbers, a 95% confidence interval, and a brief note on key assumptions. Also list 2 data checks to validate before trusting the forecast."

  1. Build queries and validation checks

"Context: attribute user_tier values are free, pro, enterprise. Task: Give an Amplitude (or SQL) query to compute 7-day conversion rate by tier, and a validation checklist to ensure the data mapping is correct."

Helpful practices (do's) and small caveats

Do:

  • Ask creative, hard questions — the agent is built to explore and explain.
  • Request chart types when a visual helps (line, bar, stacked, heatmap, funnel). If a chart doesn't make sense for the requested insight, the agent will say so.
  • Ask for forecasts, scenarios (what-if), or action plans — you can get both numbers and suggested next steps.
  • Provide the names of important attributes, events, or view IDs if you want runnable code/query output immediately.
  • Iterate: start broad, then ask follow-ups that narrow or test hypotheses.

Small caveats:

  • If you don't provide attribute/event definitions, the assistant may propose reasonable guesses; verify proposed column names before running suggested queries.
  • Forecasts are only as good as the input data; ask for assumptions and validation checks.

Advanced tips for data-savvy prompts

  • Give example values for categorical fields so filters are exact (e.g., country = US, CA, GB).
  • Ask for smoothing/windows explicitly: "7-day smoothing" or "week-over-week comparison".
  • When asking for code/API payloads, request inline comments or a brief explanation for each step.
  • Ask the assistant to produce a short validation checklist (what to inspect after applying a fix).

Sample prompt snippets you can paste:

"Show revenue as a line chart, 7-day moving average, last 90 days. Then break down the last 14 days by product_category as a bar chart."

"Forecast next quarter revenue with a simple exponential smoothing model; show forecast and 80/95% intervals and list 3 assumptions."

"Run a root-cause analysis for the decline in desktop conversions vs mobile: give top 3 hypotheses and one query or chart to validate each."

Use this guide as a quick reference when you click "Ask the agent." The assistant will often handle vague questions well — but giving it attributes, events, and views from your Anamap project will make its answers faster, more accurate, and directly runnable. If you'd like, we can add small prompt templates to the UI that insert these examples into the chat input.