13 best Claude skills for data analysts (Shopify, ads, and returns exports)

Ask Claude "is this product profitable?" with a CSV pasted in and you'll get a confident answer built on assumptions the model invented and didn't tell you about. The number will look reasonable. You won't be able to check it, and you won't know which costs it silently left out. Banning "just estimate it" in the prompt moves the problem somewhere else - the model still fills the gap, it just phrases the guess more carefully.

The best Claude skills for data analysts each carry a stated method - a defined margin stack, a defined days-of-cover formula, a defined dispute-dating rule - so the output shows its work and flags what's missing instead of guessing a value to fill the gap. That's the pack in this post. Thirteen skills, MIT-licensed, ordered so you know which one earns its install first.

For context on why margin skills matter more than dashboard skills right now, McKinsey's 2024 State of AI report found analytics and finance teams are the fastest-growing users of generative AI inside enterprises, and Salesforce's 2024 State of Data & Analytics reported 41 percent of analysts say inconsistent data definitions - not tooling - are their biggest blocker to trustworthy analysis. A skill with a stated method fixes the definition problem at the paste boundary. For the positioning read on this role, see the data analyst is the GTM job AI actually replaces. For a broader tour of the whole catalog, see best analytics tools for GTM.

What is a Claude skill

A Claude skill is a versioned SKILL.md file with a stated method: the inputs it asks for, the framework it applies, and the shape of the output it returns. You install a pack once with npx skills add sidchaudhary/gtm-skills, restart Claude Code, and every skill in the folder is available by name. Two analysts running the same skill get the same output shape - a CM1/CM2/CM3 table, a banded shipping recovery report, a dated chargeback cluster - so the work stops drifting between sessions and between people.

Why data analyst teams need Claude skills now

Analysts spend most of their week on the shaping layer - pulling exports, cleaning columns, joining them by hand - and shrink the actual analysis to whatever's left on Friday. Salesforce's 2024 State of Data & Analytics reports 41 percent of analysts name inconsistent definitions as their top blocker, and HubSpot's 2024 State of Marketing report puts reporting and measurement in the top three time drains for revenue teams every year running.

Claude skills fix this at the task level. You do not replace the analyst. You give the analyst a stated margin formula, a stated days-of-cover rule, and a weekly readout skill that reads the outputs of the deep skills instead of re-deriving them, so the strategic layer stops taking a full week per question.

The skills that move the needle are not the ones that generate a chart. They're the ones that answer three questions faster:

• Which SKUs actually make money after every variable cost, not the blended margin the dashboard shows?
• Which risks - stockouts, returns, disputes - would make the margin number a lie the moment you act on it?
• What single readout captures what changed this week, without asking anyone to open five dashboards?

Install the pack in about 30 seconds

One command pulls all 13 Data Analyst skills into your project's .claude/skills/ folder. Restart Claude Code afterward - skills load at startup, so a running session won't see them.

npx skills add sidchaudhary/gtm-skills/skills/data-analyst

# restart Claude Code, then confirm:
# "What skills do you have for analysis?"

You can install a single skill instead by appending its folder name, e.g. .../skills/data-analyst/the-margin-builder. The pack install is the better default because these skills chain - the-weekly-reporter is meant to read the-margin-builder's output as an input, and skipping the second means the first re-derives a shallower number every Monday.

13 best Claude skills for data analysts

1. the-margin-builder

Builds a per-SKU or per-order contribution margin stack - CM1 after product cost, CM2 after fees and shipping, CM3 after allocated ad spend - from raw revenue, COGS, fee, and spend inputs. If an input is missing it asks. Anything still missing gets marked as a stated assumption rather than filled with a guess, which is the detail that makes the output actually trustworthy on a low-margin SKU.

When to use: Any time a margin question comes up and the honest answer so far has been "about 40 percent, I think." Especially before scaling ad spend on a SKU that looks fine on ROAS but might be underwater at CM3.

How to run:

You: "Run the-margin-builder on this order export."
      [paste: order id, revenue, COGS, fees, shipping charged, ad spend]

Claude: builds the CM stack per SKU
        -> CM1 after product cost
        -> CM2 after fees + shipping
        -> CM3 after allocated ad spend
        -> flags every input it did not receive as a stated assumption

Best used with: An order-level export that includes COGS. If COGS lives in a separate sheet, join them first - the skill will tell you it's missing rather than invent a number, which is the correct behavior but not the fastest path.

Where this skill hits its ceiling: It doesn't connect to your database. It analyzes what you paste. If your real bottleneck is "someone has to pull the export," this skill doesn't remove that step.

With Intempt: Intempt's Data Analyst agent runs the same CM stack against live event data instead of pasted exports, so the number updates when a supplier price does, not when someone remembers to re-run it.

Get the-margin-builder →

2. the-shipping-recovery-check

Compares shipping charged against shipping paid, banded by order value and zone, and tests whether your free-shipping threshold is actually pulling weight or just leaking margin. The banding is the important part - a blended shipping recovery ratio hides that small orders are subsidizing large ones, or that one zone eats all your losses.

When to use: When small orders look worse than large ones on the margin stack, or when carrier rates just moved and nobody has re-checked the threshold in six months.

How to run:

You: "Run the-shipping-recovery-check on this order + carrier export."
      [paste: order value, zone, shipping charged, actual carrier cost]

Claude: -> recovery ratio by order value band
        -> recovery ratio by shipping zone
        -> where free-shipping threshold sits vs median AOV
        -> whether the threshold is actually moving order size or not

Best used with: Actual carrier invoices, not a rate-card estimate. The gap between the two is usually where the leak lives.

Where this skill hits its ceiling: It checks recovery, not carrier selection. If the answer is "switch carriers," that's a procurement decision, not one the skill makes for you.

With Intempt: Shipping recovery is scored continuously against live order data, so a threshold that drifts underwater is flagged before the next quarterly review.

Get the-shipping-recovery-check →

3. the-inventory-risk-scanner

Scores stockout and overstock risk at the SKU level from sales velocity, on-hand units, and lead time, using a stated days-of-cover threshold instead of a gut read. The output is a ranked list of SKUs at risk with a named reason - not "low stock," but "14 days of cover against a 21-day lead time on your top-decile revenue SKU."

When to use: Before scaling ad spend on a SKU (can you even supply it), and any time inventory has grown past what someone can spot-check by eye. Also useful as a monthly sweep - the overstock side of the report catches SKUs bleeding storage costs while nobody's looking.

How to run:

You: "Run the-inventory-risk-scanner on this SKU export."
      [paste: SKU, on-hand units, 30-day sales, lead time]

Claude: -> days of cover per SKU against stated threshold
        -> stockout risk, ranked by revenue at stake
        -> overstock risk, ranked by units and storage load
        -> the named reason for each flag

Best used with: A recent sales velocity window, not a trailing year. Seasonal SKUs blow up the average and a rolling 30 or 60 days gives a truer signal.

Where this skill hits its ceiling: It scores risk, not the reorder quantity. Deciding how many units to reorder is a purchasing conversation the skill doesn't try to run.

With Intempt: Inventory risk is joined to ad spend automatically, so you don't pay to advertise a product you can't ship - that's the-stockout-spend-guard loop in the Store Loops pack.

Get the-inventory-risk-scanner →

4. the-returns-miner

Groups structured return reason codes to specific SKUs and surfaces the highest-volume root causes behind returns. The point isn't the return rate - it's whether returns concentrate around three SKUs with a coded reason ("sizing runs small," "photo doesn't match color") or scatter randomly across the catalog. Concentration is a fix. Scatter is a policy conversation.

When to use: When return rates are rising and the reason codes exist but nobody has grouped them. Also useful right after a photography or PDP refresh - the return codes will tell you within four weeks whether the refresh actually helped.

How to run:

You: "Run the-returns-miner on this RMA export."
      [paste: SKU, return date, coded reason, order value]

Claude: -> SKUs with the highest return concentration
        -> reason code driving each one
        -> revenue and margin impact per SKU
        -> which returns look like a PDP fix vs a product fix

Best used with: Structured reason codes, not free-text notes. If your returns data is unstructured, use the-theme-miner instead - it's built for messy text and will get you to the same insight from a different input.

Where this skill hits its ceiling: It groups the codes you have. If your reason codes are three options and everyone picks "other," no skill fixes that - the return form does.

With Intempt: Return concentration feeds straight into the PDP audit and the product roadmap, so the same signal doesn't get re-discovered every quarter.

Get the-returns-miner →

5. the-chargeback-pattern-finder

Clusters chargebacks by product, channel, geography, order value band, and payment method to separate real fraud from service failure and friendly fraud. Dates every dispute to the original sale rather than the filing date, which is the detail a manual read almost always gets wrong - and the one that decides whether the pattern is your fraud rules or a policy change you made three weeks ago.

When to use: When dispute rates are climbing and the honest answer to "is it fraud?" has been a shrug. Also worth running after any change to fulfillment SLAs, shipping partners, or return policy - service failure disputes show up 30 to 60 days later, on the sale date the skill uses, not the filing date the report shows.

How to run:

You: "Run the-chargeback-pattern-finder on this dispute export."
      [paste: dispute id, sale date, filing date, reason code,
              product, channel, geo, order value, payment method]

Claude: -> disputes clustered by pattern, dated to the sale
        -> real fraud vs service failure vs friendly fraud
        -> the change that likely caused each cluster
        -> ranked by dollars, not count

Best used with: A dispute export that includes both sale date and filing date. If you only have filing date, the skill will still run but it will flag that the clustering may misdate the cause.

Where this skill hits its ceiling: It names the cluster. Fixing a service-failure cluster is an ops conversation the skill doesn't have.

With Intempt: Dispute patterns are joined to the customer journey, so a friendly-fraud cluster from a single channel is caught in the segmentation, not just the finance report.

Get the-chargeback-pattern-finder →

6. the-cohort-tracker

Groups customers by acquisition period and tracks a metric across the periods that follow - retention, revenue, order frequency, whatever you point it at. Blended averages hide the fact that newer customers might be worse than older ones, or that one acquisition month is carrying the whole LTV curve. Now also computes a CAC payback window per cohort when acquisition spend is supplied.

When to use: Monthly, once your customer base is old enough that cohorts differ. Also useful before any large channel shift - if you scaled Meta in Q2, the Q2 cohort will tell you within 90 days whether that traffic actually retained.

How to run:

You: "Run the-cohort-tracker on this customer export.
      Metric: repeat revenue per customer."
      [paste: customer id, first-order date, subsequent order dates + revenue]

Claude: -> cohorts grouped by acquisition month
        -> retention and revenue curve per cohort
        -> which cohorts diverge from the blended average
        -> CAC payback window per cohort (if spend supplied)

Best used with: At least six months of order data. A three-month cohort read is a preview, not a signal.

Where this skill hits its ceiling: It reports the divergence. The reason a specific cohort is better or worse is a segmentation question the skill hands off to the-theme-miner or your CRM.

With Intempt: Cohorts run continuously against live order data and feed the lifecycle segmentation directly, so a bad cohort triggers a lifecycle response instead of a monthly review.

Get the-cohort-tracker →

7. the-weekly-reporter

Takes your performance, traffic, lifecycle, inventory, and support summaries and produces one weekly operating readout: what changed, what likely caused it, what to do next. Meant to replace five scattered dashboards, not compete with them. The move that makes it actually good is feeding it the *output* of the deep skills as inputs - run the margin stack, paste its summary in, and the weekly readout inherits a real number instead of re-deriving a shallow one.

When to use: Every Monday. Standing slot on the calendar. Every other skill in this pack earns its slot against a specific decision - this one earns it against the calendar.

How to run:

Every Monday:
  1. paste week's exports    -> the-weekly-reporter
  2. paste last margin run   -> same session, as an input
  3. paste last returns run  -> same session, as an input

Claude: -> what moved this week vs last
        -> the likely cause behind each move
        -> 3 next actions, ranked by dollars
        -> one page, not five dashboards

Best used with: The outputs of the other skills in this pack. A weekly readout run against raw exports is a summary; a weekly readout run against summarized skill outputs is a real read.

Where this skill hits its ceiling: It summarizes across whatever you bring. Don't ask it to do a deep read on one area - run that skill directly and feed its result in here.

With Intempt: The weekly readout is generated from the same live data the deep skills read, so the Monday review isn't a paste chain - it's already there when you open the tab.

Get the-weekly-reporter →

The other 6 skills in the Data Analyst pack

The seven above are the ones most analysts install first because they map directly to money questions. The remaining six sit alongside them and earn their slot against specific triggers rather than a standing cadence.

• the-kpi-blueprint: KPI dashboard specs with metric formulas, visualization types, alert thresholds, and layout wireframes. Run once when the team doesn't trust the numbers on the current dashboard.
• the-lever-finder: Prioritized growth strategy with maturity assessment, growth levers, channel priorities, and a quarterly plan. Run when the team is spread across too many bets.
• the-theme-miner: Turns transcripts, reviews, tickets, and surveys into confidence-scored themes, quote banks, and personas. The unstructured-text counterpart to the-returns-miner.
• the-anomaly-alert: Flags which recent points in a metric are genuinely outside its normal range, not a gut read of a chart. Run when a number moved and nobody's sure if it's real.
• the-benchmark-check: Checks one of your metrics against a real benchmark source with stated provenance. Run when someone asks "is this number good?" and the honest answer has been a guess.
• the-catalog-auditor: Audits product catalog completeness across the full catalog, not a sampled spot-check, and names the highest-revenue SKUs missing what.

Comparison: which Claude skill to use when

Free Intempt tools that pair with these skills

Browser-based, sit behind free-plan signup, produce quick outputs you can feed into the skills above.

• Marketing ROI calculator: First-pass ROI read on a channel or campaign. Use as a sanity check before running the-margin-builder on the underlying orders.
• CAC LTV calculator: Quick CAC and LTV numbers per channel. Feed the output into the-cohort-tracker to see if the blended CAC is hiding cohort divergence.
• Conversion rate calculator: Stage-by-stage conversion snapshot. Pairs with the-anomaly-alert when a rate moves and you're not sure if it's real.

What to look for in a Claude skill for data analysis

Not every prompt is worth wrapping in a skill file. The ones that stick share four traits.

Stated method. The skill names the formula, the threshold, the dating rule. You should be able to read the SKILL.md and know exactly which number the output represents before you paste anything in.

Missing-input honesty. A skill that guesses at a missing COGS is worse than one that refuses. The best ones ask, and if you can't supply it, they mark the output as an assumption rather than hide the gap.

Consistent output shape. Every run returns the same table shape so downstream work plugs in without reformatting - the-weekly-reporter can read the-margin-builder's output only because that output looks the same every run.

Single use case. Skills that try to do three things do none of them well. the-margin-builder builds a CM stack. the-cohort-tracker tracks cohorts. Neither pretends to be the other, and the pack is stronger for it.

Final thoughts: which skill to install first

Start with these three, in order:

• the-margin-builder: answer the money question first. Every other skill in the pack is easier to prioritize once you know which SKUs actually make money at CM3.
• the-inventory-risk-scanner: run it right after the margin stack. A profitable SKU you're about to stock out of needs a different decision than the margin table alone suggests.
• the-weekly-reporter: standing Monday slot. Feed it the outputs of the deep skills as inputs and it turns into a real readout instead of a surface summary.

For the positioning read on where this role is heading, see the data analyst is the GTM job AI actually replaces. If you want the whole 78-skill catalog rather than just the 13 for data analysts, browse the gtm-skills directory. Start for free on Intempt if you'd rather run the same jobs against live event data than paste exports by hand.