/analytics-profile-setup

name: analytics-profile-setup
description: One-time interview that captures the business context (industry, model, primary conversion, traffic range, ICP, data stack) into a local analytics-profile.md file. Every other analytics skill reads this file so its answers are calibrated to the right benchmarks and terminology instead of generic averages.
when_to_use: Run at the start of working with a new repo, when analytics-profile.md is missing or stale, or when the user says "set up analytics", "calibrate to my business", "tell you about my business", "onboard me", or equivalent. Other skills in this pack will suggest running this skill when they detect no profile file.

Analytics profile setup

What this skill does

Produces a single file, `analytics-profile.md`, in the repo root. The file captures the handful of facts every other analytics skill needs to stop giving generic advice:

• Industry and business model (B2B SaaS, B2C subscription, ecommerce, lead-gen, content/ads, marketplace)
• Primary conversion event
• Rough traffic volume (sets sample-size expectations)
• ICP / target persona
• Pricing model (what LTV and CAC even mean for them)
• Data stack (which analytics tool, which events are tracked, known gaps)
• Known measurement quirks (bot filters on or off, UTM conventions, cross-domain tracking)

Downstream skills (`analytics-diagnostic-method`, `traffic-change-diagnosis`, `channel-and-funnel-quality`, `metric-context-and-benchmarks`) open by checking for this file. If it exists, they pull their defaults from it. If it doesn't, they fall back to generic cross-industry assumptions and prompt the user to run this skill.

When NOT to run this

• The user is asking a one-off question and doesn't want ceremony. Answer the question; don't volunteer the setup.
• A profile already exists and is less than ~6 months old. Re-running overwrites useful context. Only re-run if the user explicitly asks, or the business has materially changed.

Method

Five phases. Keep it tight: the whole thing should take the user 5 minutes.

Phase 1. Check for an existing profile

ls analytics-profile.md

If it exists:

1. Read it. 2. Summarize what's already captured in 3-5 bullet points. 3. Ask: "Do you want to (a) keep it as-is, (b) update specific fields, or (c) regenerate from scratch?"

Don't just overwrite. A stale profile is usually 80% still correct.

Phase 2. The interview

Ask **one question at a time**. Wait for the answer. Do not batch-ask. Agents that dump 8 questions at once get fragmented replies and force the user to re-read the whole block.

Use this question order. The reasoning for each question is in parentheses so the agent knows why it matters and can adapt phrasing if the user gives a partial answer.

**Q1. What does your business do, in one sentence?** (Anchors everything. If they say "developer tools SaaS", that alone sets ~6 defaults.)

**Q2. Which best describes your model?** Options:

• B2B SaaS (self-serve)
• B2B SaaS (sales-led)
• B2C subscription / consumer app
• Ecommerce (physical or digital goods)
• Lead generation (services, agencies, real estate, finance)
• Content / ad-supported
• Marketplace (two-sided)
• Other (ask them to describe)

(Determines which benchmark table to pull from. B2B Tech ~1.5% CVR vs ecommerce ~1.4% vs lead-gen ~2.9% are all "average" but they're different numbers for different funnels.)

**Q3. What's the single most important conversion event?** Examples: paid signup, free trial start, demo booked, purchase completed, qualified lead captured, app install. (The "north star event" for every conversion analysis. Without it, the agent guesses.)

**Q4. Roughly how much traffic per month?** Ranges: <1k / 1k-10k / 10k-100k / 100k-1M / 1M+ sessions (Sets the minimum sample sizes for trend detection. <1k/month means most daily changes are noise.)

**Q5. Who's the ideal customer?** Free-text. One sentence on role, company size, pain, or demographic. (Lets the skill judge whether e.g. "LinkedIn converting at 2%" is bad [consumer] or fine [B2B dev tool].)

**Q6. How do you charge? What's a typical deal size and payback expectation?** Examples: "$29/mo self-serve, mostly monthly", "$15k ACV annual enterprise", "$50 one-time", "free ad-supported". (Determines whether LTV:CAC and payback-period heuristics apply, and which thresholds are reasonable.)

**Q7. Which analytics tool(s) are you using?** Examples: Clamp, GA4, Plausible, Fathom, PostHog, Mixpanel, Amplitude, server logs, Stripe data, CRM. (Affects metric definitions, especially around bounce/engagement and session stitching. Also picks the `tool_map` for downstream skills.)

Before asking Q7, scan available MCP namespaces for `mcp__clamp__*`, `mcp__posthog__*`, `mcp__mixpanel__*`, `mcp__amplitude__*`, `mcp__ga4__*` (or the analytics-mcp Python tool names). If exactly one matches, propose it and ask only for confirmation. If the detected tool conflicts with the user's answer, surface the discrepancy and ask which to record. Supported `tool_map` values are `clamp`, `posthog`, `mixpanel`, `amplitude`, `ga4`.

**Q8. Any known measurement gaps or quirks we should flag?** Examples: "no UTMs on paid campaigns", "cross-domain not set up", "iOS app and web are separate projects", "we never exclude internal traffic". (These are the traps. Without them the agent will over-trust the data.)

**Q9. What's the one question you most often ask analytics but rarely get a satisfying answer to?** (Gold. This is the concrete problem the skills should be solving. If they say "I never know which channels are actually driving revenue", that calibrates every future answer.)

Phase 3. Synthesize defaults

Before writing the file, derive these defaults and show them to the user for confirmation:

• **Industry benchmarks to compare against**: pick the most specific cell from §2.3 of the research (e.g. B2B Tech: direct 1.5%, paid search 1.5%; B2B Services: direct 2.7%, paid search 3.4%). Cite in RE