ChatData
YOUR OPEN-SOURCE DATA SCIENCE WORKSPACE

Good AI answers
don’t reset.

ChatData sets up your AI’s data science workflow locally—with analysis skills, runnable checks, and saved records the next session can build on.

Free forever · Open source · 16 skills · No credit card

Usage counts, never your prompts or data. How it works ↗

At home in your workflowClaude CodeNative pluginCodexNative pluginCursorNative skills
↳ A closer look at the answer.Interactive examples · synthetic data
OVERALL CONVERSION

“Why did conversion drop?”

8%
17% → 8% · a 9 percentage-point drop
BEFOREAFTER
A second lookWhat the checks show
01Segment ratesUNCHANGED
02Customer mixSHIFTED
03ArithmeticRECONCILES

Neither segment got worse. More visitors came from the lower-converting segment. The mix change explains the full drop; it does not tell us what caused the mix to change.

Next: investigate the acquisition mix.

The bundled rate-decomposition script reproduces this example from two rows of synthetic data.

Inspect the example ↗
HOW GOOD ANSWERS CARRY FORWARD

The setup behind
the answer.

You could build it yourself: write Markdown skill files, package them for each AI client, wire up Claude Code hooks, add calculation scripts, and design records for the next session to reuse. ChatData brings those pieces together for one person.

Inspect what you install ↗

01

Install the working setup.

Get 16 analysis skills, runnable Python checks, examples, and a reusable record template. Claude Code also gets startup hooks and linked usage reporting; Codex and Cursor use their own skill setup.

02

Save the reasoning with the result.

The skills ask your agent to save the definition, source, code, checks, and caveats in a local folder you choose. You can inspect and edit the files.

03

Give the next session a starting point.

Point your AI at the saved record. It can pick up the work, rerun the calculation, and check what changed before reusing the answer. Your files carry the context between sessions.

16 SKILLS / ALL INCLUDED

For the work
on your desk.

You might be a data scientist with more questions than hours. You might be an analyst trying a new method, or the person who has to figure out the data alone. Pick the job in front of you. Each skill gives your AI a method, checks to run, and a useful result to leave behind.

Read the full capability guide ↗

Provide the relevant files or an authorized connection, plus the decision you need to make. The skills guide your AI; they use the tools available in your session. Open a workflow for its inputs, checks, output, and a prompt you can copy.

01Design a test worth running

Set the hypothesis, sample size, guardrails, and stopping rule before looking for a winner.

Bring this

Hypothesis, planned action, smallest useful effect, primary metric, randomization unit, eligibility, exposure rule, baseline, traffic, and risk guardrails.

What the skill asks your AI to do

  • Estimate sample size with assumptions.
  • Predeclare alpha, power, allocation, duration, exclusions, guardrails, and stopping rule.
  • Check interference, contamination, carryover, novelty, seasonality, instrumentation, and outcome maturity.
  • Recommend an A/A or logging check when exposure is uncertain.

You leave with: an experiment plan and power calculation.

TRY ASKING

Use ChatData’s experiment-design skill. Plan an A/B test for our signup flow.

Know the limits

The bundled power helper covers equal-allocation independent binary outcomes. Continuous, repeated, or clustered outcomes need another variance model or simulation. The skill does not launch experiments.

Read the skill ↗
02Vet an experiment

Check assignment, uncertainty, outcome maturity, and whether the effect is large enough to matter.

Bring this

Experiment design, assignment and exposure counts, outcomes, planned allocation, smallest meaningful effect, observation horizon, exclusions, and guardrails.

What the skill asks your AI to do

  • Test sample-ratio mismatch before interpreting lift.
  • Report absolute and relative lift with uncertainty.
  • Inspect maturity, missingness, attrition, leakage, delayed outcomes, invariant metrics, and guardrails.
  • Separate statistical evidence from an effect large enough to act on.
  • Treat exploratory segments as exploratory.

You leave with: a readout with a conditional decision.

TRY ASKING

Use ChatData’s experiment-analysis skill. Is this A/B test ready to call?

Know the limits

The bundled helper covers aggregate binary outcomes in two independent groups. Repeated users, clusters, paired data, heavy-tailed revenue, and continuous outcomes need a suitable estimator.

Read the skill ↗
03Find the drop-off

Count people through ordered steps. Keep duplicates, early events, and immature cohorts from distorting conversion.

Bring this

Entity identifier, entry event, ordered steps, timestamps with timezone, entry cohort, conversion window, and open or closed funnel definition.

What the skill asks your AI to do

  • Count entities rather than raw events.
  • Enforce strict step order.
  • Exclude cohorts without a fully elapsed conversion window.
  • Inspect duplicates, late arrival, identity stitching, bots, and bad event names.
  • Show entry-to-step and adjacent-step conversion with counts.

You leave with: a funnel table with checked denominators.

TRY ASKING

Use ChatData’s funnel-analysis skill. Where are people dropping out of signup?

Know the limits

The bundled helper implements a closed CSV funnel using earliest entry and strict order. Re-entry, session, equal-time, or open-funnel rules need an explicit alternate implementation.

Read the skill ↗
04Explain a metric change

Separate changes in customer mix from performance within a segment. Test alternatives before claiming a cause.

Bring this

A changed metric, comparable before and after periods, its definition, source and cutoff, and grounded segments.

What the skill asks your AI to do

  • Reproduce both values from the same definition and source.
  • Rule out partial periods, late data, event renames, duplicate joins, and source changes.
  • Separate population mix from within-segment performance.
  • Reconcile contributions to the total change.
  • Try to disprove the leading explanation.

You leave with: a reconciled breakdown and next tests.

TRY ASKING

Use ChatData’s root-cause skill. Why did conversion drop last week?

Know the limits

A decomposition explains where arithmetic movement occurred, not why behavior changed. Segment inputs must be disjoint and exhaustive; zero-denominator cells need explicit treatment.

Read the skill ↗
05Understand who comes back

Compare cohorts at the same age. Separate churn from customers who have not had time to return.

Bring this

Cohort entry, entity, return event, interval, timezone, eligibility, and the chosen retention or churn definition.

What the skill asks your AI to do

  • Build an entity-period spine.
  • Keep cohort age separate from calendar time.
  • Mask cells without enough follow-up instead of treating them as zero.
  • Compare cohorts at equal age with counts and rates.
  • Inspect resurrection, delayed activation, migration, censoring, and data coverage.

You leave with: a cohort analysis with incomplete periods marked.

TRY ASKING

Use ChatData’s retention skill. Compare retention across signup cohorts.

Know the limits

The skill does not infer lifetime value from a few early points. Survival estimates require explicit event and censoring rules; competing risks may require another estimator.

Read the skill ↗
06Check the data first

Inspect grain, missing values, freshness, duplicates, and joins against the question you need to answer.

Bring this

Dataset or schema, intended analysis, row grain, expected keys, time fields, coverage, and tolerance for error.

What the skill asks your AI to do

  • Test uniqueness, missingness, types, ranges, drift, integrity, time bounds, and impossible sequences.
  • Record table grain and key cardinality before joins.
  • Reconcile rows, entities, amounts, and unmatched keys after joins.
  • Quantify each issue's count, share, and likely impact.
  • Return usable, usable with limitations, or blocked for the named question.

You leave with: a suitability assessment and specific fixes.

TRY ASKING

Use ChatData’s data-quality skill. Can I trust this CSV for the analysis?

Know the limits

The bundled CSV profiler checks rows, missing values, duplicate rows, and duplicate keys. Those checks alone do not prove freshness, business validity, representativeness, or join safety.

Read the skill ↗
07Review the query

Catch join inflation, wrong time windows, hidden exclusions, and averages with the wrong weights.

Bring this

SQL dialect, query, schemas, source and result grain, key relationships, timezone, and metric definition.

What the skill asks your AI to do

  • Review join cardinality and preaggregate many-side tables.
  • Check NULLs, counts, DISTINCT, LEFT JOIN filters, integer division, zero denominators, and ratio aggregation.
  • Use explicit timezone and half-open time ranges.
  • Check windows for partition, order, tie breakers, and frame.
  • Validate boundary fixtures and reconcile an independent total.

You leave with: reviewed sql and boundary-case checks.

TRY ASKING

Use ChatData’s sql-review skill. Review this query before I use the number.

Know the limits

Review does not prove execution. Scan cost is reported only when observed; LIMIT can reduce returned rows without reducing bytes scanned.

Read the skill ↗
08Find useful questions

Explore distributions, outliers, and segments while keeping promising patterns separate from confirmed findings.

Bring this

An unfamiliar dataset, its collection process and coverage, and the decision exploration should inform.

What the skill asks your AI to do

  • Start with grain, data quality, distributions, missingness, outliers, and segment counts.
  • Use robust summaries for skewed data.
  • Track comparisons tried and separate exploration from confirmation.
  • Check time trends, confounding, selection, Simpson's paradox, nonlinearity, and leakage.
  • Return a data map and a short ranked list of testable questions.

You leave with: an exploration with testable next questions.

TRY ASKING

Use ChatData’s exploratory-analysis skill. What is worth investigating in this dataset?

Know the limits

Clusters, embeddings, surprising slices, and exploratory associations are hypotheses, not discovered truth or causal evidence.

Read the skill ↗
09Forecast with a baseline

Use time-based backtests, compare simple baselines, and show uncertainty at the horizon that matters.

Bring this

Target series, dated observations, forecast horizon, decision cadence, known future inputs, and costs of over- and under-prediction.

What the skill asks your AI to do

  • Inspect gaps, seasonality, structural breaks, exposure changes, and zeros.
  • Keep a temporal holdout.
  • Compare naive and seasonal-naive baselines with rolling-origin evaluation.
  • Fit preprocessing on training data only.
  • Evaluate uncertainty coverage and width by horizon.

You leave with: a forecast with backtests and limits.

TRY ASKING

Use ChatData’s forecasting skill. Forecast demand for the next four weeks.

Know the limits

In-sample fit is not forecast validation. Future features must be available at prediction time, and regime changes may support scenario bounds rather than precise intervals.

Read the skill ↗
10Build a model you can assess

Define the prediction moment, prevent leakage, and inspect calibration and errors across relevant groups.

Bring this

Prediction moment, label and delay, unit, action, error costs, available-at-prediction features, and historical data.

What the skill asks your AI to do

  • Choose time or group splits that match production.
  • Fit preprocessing only on training data.
  • Compare a simple baseline.
  • Check actionable thresholds, calibration, subgroup errors, residuals, and shift.
  • Search for post-outcome fields, future aggregates, duplicated entities, and other leakage.

You leave with: a baseline comparison and model card.

TRY ASKING

Use ChatData’s predictive-modeling skill. Build a churn model and check for leakage.

Know the limits

Feature importance explains model behavior rather than causal effects. Deployment or automated decisions about people require separate authorization and domain review.

Read the skill ↗
11Question the causal claim

Name the assumptions behind an intervention effect. Check overlap, timing, and alternative explanations.

Bring this

Treatment, outcome, target population, assignment mechanism, time zero, follow-up, estimand, and an explicit causal structure.

What the skill asks your AI to do

  • Name confounders, mediators, colliders, and post-treatment variables.
  • State the identification assumptions for the selected design.
  • Check overlap, balance, pretrends, manipulation, timing, or instrument assumptions as relevant.
  • Use negative controls, placebos, sensitivity analysis, and alternative specifications.
  • Separate descriptive association from a supported intervention effect.

You leave with: an estimate with identification checks.

TRY ASKING

Use ChatData’s causal-inference skill. Did the launch cause this change?

Know the limits

Passing observed diagnostics cannot prove untestable causal assumptions. If the design does not identify the effect, the skill should return a descriptive result and an evidence plan.

Read the skill ↗
12Make the result readable

Choose the right comparison, label the units, show uncertainty, and verify the rendered chart against the data.

Bring this

The reader's decision, checked calculation, source and refresh date, intended display size, and available charting tools.

What the skill asks your AI to do

  • Choose an encoding that matches the comparison.
  • Label units, dates, population, source, sample size, uncertainty, and missing periods.
  • Compare like periods, scales, and denominators.
  • Verify plotted values against the calculation.
  • Inspect the rendered output for clipping, legends, axes, and accessible contrast.

You leave with: a chart or dashboard with source notes.

TRY ASKING

Use ChatData’s visualization skill. Turn this analysis into a chart I can share.

Know the limits

The skill uses the charting tools already available to the AI client. It does not publish or upload an artifact without authorization.

Read the skill ↗
13Get the definition right

Specify the population, numerator, denominator, exclusions, timezone, and source before calculating a KPI.

Bring this

The business decision, existing metric definitions or queries, source schema, and boundary cases.

What the skill asks your AI to do

  • Specify unit, numerator, denominator, eligibility, exclusions, timezone, window, aggregation, source, and delay.
  • Aggregate numerator and denominator before dividing.
  • Test cases such as refunds, duplicates, returning users, and late records.
  • Show competing definitions when ambiguity changes the number.

You leave with: a metric definition with boundary examples.

TRY ASKING

Use ChatData’s metric-definition skill. Define activation for this product.

Know the limits

It cannot make a metric canonical without the owner's approval or calculate a metric the supplied data cannot support.

Read the skill ↗
14Pressure-test the answer

Recompute a critical result, look for a plausible failure, and check whether the conclusion goes beyond the evidence.

Bring this

The actual analysis inputs, code or query, output, definition, and any saved analysis record.

What the skill asks your AI to do

  • Recompute at least one decision-critical result when access allows.
  • Check grain, denominator, maturity, freshness, joins, missingness, selection, multiplicity, assumptions, leakage, and causal language.
  • Test a plausible boundary case.
  • Look for sensitivity that reverses the conclusion.
  • Return supported, supported with limitations, or not supported, then fix and retest straightforward local defects.

You leave with: a review with corrections and unrun checks.

TRY ASKING

Use ChatData’s analysis-review skill. Try to disprove this analysis.

Know the limits

A review can only verify evidence it can inspect. Missing reproducibility blocks a verified-reuse claim; it does not justify inventing a source path or pretending a check ran.

Read the skill ↗
15Turn evidence into a decision

Explain the recommendation, the strongest objection, and what new evidence would change the next step.

Bring this

A decision, analysis and source record, supported alternatives, uncertainty, timing, and known owner or deadline.

What the skill asks your AI to do

  • Lead with the recommended decision and confidence.
  • Separate observed facts, estimates, assumptions, and hypotheses.
  • Include denominators and periods for important numbers.
  • Show the strongest evidence against the recommendation.
  • Tie the next action to evidence that would change the decision.

You leave with: a short brief linked to the calculation.

TRY ASKING

Use ChatData’s decision-brief skill. What should we do based on this analysis?

Know the limits

The skill will not invent urgency, customers, benchmarks, savings, owners, or deadlines. Sending or publishing the brief is a separate action.

Read the skill ↗
16Start with the question

Choose the right workflow from your question and available data, then carry it through the relevant checks.

Bring this

A decision or question, the available dataset or query, and the cost of a wrong answer.

What the skill asks your AI to do

  • Inspect supplied files before asking for details already present.
  • Route to one relevant specialist skill.
  • Run the calculation and relevant checks instead of stopping at a plan.
  • Label bundled examples as synthetic.
  • Save a local analysis record when the result should be reused.

You leave with: a completed analysis with a reusable record.

TRY ASKING

Use ChatData’s data-science skill. Help me choose the right analysis for this question.

Know the limits

The routing skill does not create data access. The local setup doctor verifies three bundled calculations, not client discovery, live sources, or every future model answer.

Read the skill ↗
WHAT CHATDATA ADDS

The code is
part of the job.

Your AI can already write SQL and make a chart. ChatData adds a consistent way to choose the method, find the mistakes that still look plausible, and keep the evidence behind the answer.

The taskWhat the skill asks your AI to check
Call an experimentAssignment balance, effect size, uncertainty, stopping rules, and guardrails before recommending a decision.
Explain a dropWhether the measurement changed, whether the customer mix shifted, and what evidence would disprove the leading explanation.
Build a modelWhat was knowable at prediction time, whether the split leaks information, and whether a simple baseline does as well.
Reuse an answerThe definition, source, calculation, checks, and caveats in a local record you can inspect and rerun.

Includes runnable Python checks for binary experiments, sample size, ordered funnels, rate decomposition, and CSV quality. Other workflows use the code and authorized tools available in your client. See what we test ↗

Keep the work for the next run. For reusable analyses, the skills ask your agent to save a local record with the definition, source, code, checks, and caveats. Point the next session at that record, then recheck what has changed. Start with a folder you choose: ChatData inventories analysis files modified in the last 30 days and creates a visible local context folder. Your originals stay in place. Add reviewed definitions and evidence before reusing a metric. See the analysis record ↗

See what your AI knows.
And what it still needs.

Your definitions should survive the next chat. Keep definitions, source checks, and review notes in your own project, with files you can read, review, and keep.

Find your recent work

Choose one project folder. The local onboarding report lists analysis files modified in the last 30 days, skipped locations, and incomplete coverage. It uses file metadata, not a hidden scan of your conversations. It never moves your originals.

Give metrics a shared format

Keep datasets, fields, relationships, and metric expressions in an Apache Ossie semantic model. ChatData includes a pinned draft schema and local validation; your evidence and review notes live alongside it. Nothing is automatically approved.

Stop when the evidence fails

Before a trusted metric answer, check its reviewed definition, source fingerprints, freshness, and validation evidence. Missing definitions, changed files, stale evidence, or unresolved conflicts block reuse and show what needs fixing.

The local check verifies recorded prerequisites. It does not prove the analysis is correct or control every prompt in your AI client. See the local workflow and its limits ↗

Your project.
Your working files.

Install the skills where you work. The bundled Python helpers read local inputs and run calculations without network requests. ChatData does not receive your datasets, prompts, results, or saved analysis records through this plugin.

What lives with you

The installed skills, example CSVs, calculation scripts, and the analysis records your agent writes to your chosen folder. Your analysis stays in those files. Linking the plugin to your personal dashboard sends usage metadata, never the content of your work.

What your AI client handles

Claude Code, Codex, or Cursor may send prompts, file contents, and tool output to its model provider. Connected data tools and any cloud sessions have their own data handling. Local plugin files do not make cloud inference private or offline.

See how you use it

Your private dashboard counts tracked ChatData workflows and their elapsed time. It starts with an editable 30-minute workflow and $125/hour to estimate time and analysis cost saved.

You choose what your agent can read, execute, save, or share. Keep sensitive files out of public repositories and follow the data permissions that apply to your work. Read the privacy policy ↗

YOUR OWN AI DATA WORKBENCH

Free now.
Free forever.

One account for your installation steps, tracked usage, and estimated time saved. Every skill is included.

Get your free plugin ↗

Already signed up? Open your dashboard ↗

A few practical details.

How does ChatData keep good answers from resetting?

The skills tell your agent to leave a local analysis record: the question, metric definition, source, calculation, checks, caveats, and what would make the answer out of date. Keep that folder and point your next session at it. The agent can read the record and recheck the inputs before continuing. Reuse depends on saving the record and following that workflow; installing the plugin alone does not make every chat remember everything. See the exact record template.

What does “free forever” include?

The ChatData plugin, all 16 skills, the example data, and the bundled analysis tools. ChatData code is open source under the MIT license; the bundled Apache Ossie schema carries its Apache license, with no trial, license key, or paid feature unlock. Your chosen AI client, model usage, and data services may charge their own fees.

Does ChatData need access to my data?

ChatData does not receive your data through the plugin. The calculation helpers run locally. Your email-linked installation sends counts, skill names, timestamps, and elapsed time. Your AI client can still send prompts, file contents, and tool output to its model provider. Read what stays local and what other services handle before using sensitive data.

Can it run experiments for me?

It can help design and vet experiments, calculate sample size, analyze results, and run local calculations or simulations through your client. Launching a production experiment or changing traffic requires your explicit approval and the appropriate tools.

Can I trust every answer?

You should review decisions that matter. The skills tell your AI to check assumptions, retain evidence, and flag missing information; they cannot guarantee that a model will follow every instruction. The repository includes executable checks and review cases so you can inspect the work.