Stats
Actions
Tags
From noibu
Analyze segment and traffic performance using Noibu data. Use when you want to know which channels, devices, or countries convert best, how mobile compares to desktop, where your traffic is coming from, which segments are underperforming, or where to find your best and worst converting customer segments.
How this skill is triggered — by the user, by Claude, or both
Slash command
/noibu:segment-analysisThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
This skill has two entry points. Read the prompt carefully before doing anything.
This skill has two entry points. Read the prompt carefully before doing anything.
If the user asked a specific, focused question ("which channel converts best?", "how does mobile compare to desktop?", "which countries are underperforming?"):
Run only the one or two queries needed to answer it.
Give a short, direct answer — a few sentences and a small table if useful.
Use AskUserQuestion — not prose — to offer the full analysis. The tool
renders a choice UI the user can act on; prose questions get skipped.
Question: "Want me to run a full segment analysis?" Options:
If they select yes, proceed to Full analysis.
If the user asked for a broad analysis or said yes to the quick-answer offer, run the two-part workflow below.
Tell the user what's happening before queries run — something like: "Starting with a high-level look across your key segments — device, country, and marketing channel — to see where conversion stands and spot anything worth investigating further."
Fire all three queries in a single turn — do not wait for one before launching the next. Also resolve any field name ambiguities in this same turn if you haven't already.
Do not apply minimum session thresholds at this stage. See the full picture first; you'll calibrate thresholds when running follow-ups.
Note on field discovery: The descriptions below explain what each query should measure conceptually. Use the Noibu connector's schema or context tools to confirm the current field names before running. Do not guess field names.
Query session data grouped by the field representing device type. Measure session count, conversion rate, and revenue per session. Order by conversion rate descending.
Query session data grouped by the field representing country (limit 30). Measure session count, conversion rate, and revenue per session. Order by sessions descending.
Order by sessions (not CVR) so the highest-traffic markets appear first — easier to spot high-volume underperformers.
Query session data grouped by the fields representing UTM source and UTM medium together (limit 30). Measure session count, conversion rate, and revenue per session. Order by sessions descending.
Same rationale — order by sessions to surface the channels carrying the most weight, then assess their conversion rate.
After the queries return, do not immediately render a full dashboard. Compute the site-wide conversion rate from this data (total sessions ÷ total purchases across all segments) and use it as the benchmark for calling out segments that are notably over- or under-performing.
Identify the 2–4 most interesting signals using the diagnostic playbook below. Follow-up queries are not predetermined — they depend on what the data shows.
| If you see this... | Consider this follow-up |
|---|---|
| A high-traffic country with conversion well below the site average | Break down by landing page — look for a localized or campaign-specific page that might have friction |
| A country with near-zero conversion | Break down by landing page filtered to that country, or check if a campaign is sending traffic to a broken page |
| A paid social channel with high traffic share but very low conversion | Break down by individual campaign to find which ones are dragging the channel down |
| One channel converting 5–10x better than others | Go deeper on that channel — by campaign or source — to find opportunities to scale it |
| Mobile conversion notably lower than desktop (or vice versa) | Break down by funnel stage and device — to find where specifically shoppers are dropping off |
| A large share of sessions with no channel attribution | Look at landing pages or referral sources to understand what's driving untagged traffic |
| Revenue per session and conversion rate moving in opposite directions for a segment | Flag as likely a currency or order value difference — may not need a follow-up query |
This step is mandatory — always perform it without exception. Do not skip or summarize it away.
Write a short, plain-language message that:
It should read like a colleague saying: "Canada's conversion is at 0.3% versus your 2.1% site average, so I'm going to look at which landing pages that traffic is hitting — that's usually where the friction shows up." Not: "Proceeding to Phase 2 deeper analysis."
Then either:
Run only the follow-up queries identified above — not a fixed set.
Apply traffic thresholds now, calibrated to this store's volume:
Campaign breakdown — when one channel's aggregate conversion rate hides variation between individual campaigns within it. Query session data grouped by UTM campaign and UTM medium together (limit 25). Order by sessions descending. Apply the same measures as the broad overview (session count, CVR, revenue per session).
Landing page breakdown — when a country or campaign seems to have friction at the entry point. Query session data grouped by the landing URL field (limit 25). Order by sessions descending. Filter to the relevant country or campaign from the broad overview to keep results focused.
Funnel stage breakdown — when you want to know where in the checkout a segment drops off. Query session data grouped by funnel depth and one additional dimension (device type, country, or UTM medium — whichever the overview flagged). Order by sessions descending. Choose the second dimension based on what the overview showed; computing step-to-step rates by segment reveals exactly where friction is highest.
Guiding principle: insights first, data second. The report must be scannable in 30 seconds. Every table that follows is supporting evidence, not the headline.
Structure:
Use AskUserQuestion:
If they choose live dashboard:
Do NOT save the already-rendered show_widget HTML — that HTML has data baked
in and will appear empty when the artifact is reopened.
Instead, build a brand-new dynamic artifact using create_artifact. The artifact
must fetch its own data every time it opens.
Important: Use the exact field names that were confirmed to work during the broad overview queries earlier in this session — do not hardcode or guess field names. The correct names will already be known from the successful queries above.
Embed config at the top as JS constants:
const DOMAIN_ID = "...";
const START_TIME = "...";
const END_TIME = "...";
On page load, call window.cowork.callMcpTool() for the three broad
overview queries (device breakdown, country breakdown, channel breakdown),
plus any follow-up queries that were run during the session. Use the exact
same field names and parameters that produced results during the analysis —
copy them directly from the working queries above.
Critical implementation details:
callMcpTool() requires the fully-qualified tool name — the same
mcp__<server>__<tool> string used in the mcp_tools allowlist. The
short name will be rejected. Store it as a constant:
const SESSION_TOOL = "mcp__<server-id>__<tool-name>";
Use the exact fully-qualified name that was active during the session.
callMcpTool() returns a content-wrapped object, not a plain parsed
response. Parse records like this:
function records(res) {
try {
let obj = res;
if (typeof res === "string") obj = JSON.parse(res);
if (obj && obj.content) {
const text = Array.isArray(obj.content)
? obj.content[0].text : obj.content;
obj = typeof text === "string" ? JSON.parse(text) : text;
}
return obj.data.domain.explorationsQueryV2.records;
} catch(e) { return []; }
}
After all fetches resolve, generate the Key Findings & Recommended
Actions section dynamically by calling window.cowork.askClaude() with
all fetched data embedded directly in the prompt. Do not bake static findings
text into the artifact HTML.
const findingsPrompt = `
You are analyzing ecommerce segment data. Based on the data below, write
3–5 numbered finding + action pairs for the Key Findings & Recommended
Actions section, ordered by revenue impact. Each finding must name a
specific segment and include a concrete number. Each action must be specific
enough to hand off to a developer or marketer. No tables — prose only.
Device data: ${JSON.stringify(deviceRows)}
Country data: ${JSON.stringify(countryRows)}
Channel data: ${JSON.stringify(channelRows)}
Follow-up data: ${JSON.stringify(followUpRows)}
Site-wide conversion rate: ${sitewideCVR}%
`;
const findings = await window.cowork.askClaude(findingsPrompt, []);
Then call askClaude() a second time to generate the per-dimension insight
callouts for the Segment Overview section, passing each dimension's rows in
the prompt.
The return value may be a string or an object — coerce it before use:
function toStr(v) {
if (typeof v === "string") return v;
if (v && typeof v === "object")
return v.text || v.content || v.message || JSON.stringify(v);
return String(v ?? "");
}
Render using the same visual structure and styling tokens as the
in-session report. The Key Findings & Recommended Actions section must be
populated from the askClaude() result — never from static text.
List all Noibu tool fully-qualified names used in the mcp_tools array of
create_artifact.
If they choose PDF:
Before invoking the pdf skill, generate a print-optimized HTML version
of the report. Web fonts (Inter, IBM Plex Mono) do not load reliably in the
PDF renderer, causing garbled or missing text.
Write this simplified HTML to a temp file, then pass it to the pdf skill.
npx claudepluginhub noibu/ai-plugin --plugin noibuCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.