Analyze a dataset and re-check every figure

---
name: analyze-a-dataset
description: "Analyzes a CSV or table: profiling, cleaning rules, answers and sanity checks run in parallel, every figure is re-checked in a fresh context, then a data report."
spec: skill-graph/1
topology:
  nodes:
    - id: goal
      type: input
      label: DATA FILE + QUESTIONS
    - id: splitter
      type: orchestrator
      label: SPLIT
      model: sonnet
      tools:
        - Read
    - id: profile
      type: worker
      label: Profile
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: cleaning
      type: worker
      label: Cleaning Rules
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: analysis
      type: worker
      label: Analysis
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: sanity
      type: worker
      label: Sanity Checks
      model: haiku
      tools:
        - Read
        - Grep
        - Glob
        - Bash
    - id: verifier
      type: verifier
      label: VERIFIER (Fresh Context)
      isolated_context: true
      model: opus
      tools:
        - Read
        - Grep
        - Glob
      gate: Every kept figure says how it was computed and matches the data on a spot check.
      on_fail: retry
      max_retries: 1
    - id: merger
      type: synthesizer
      label: FINAL ANSWER
      model: sonnet
      tools:
        - Read
  edges:
    - from: goal
      to: splitter
    - from: splitter
      to:
        - profile
        - cleaning
        - analysis
        - sanity
      contract: subtask_json
    - from:
        - profile
        - cleaning
        - analysis
        - sanity
      to: verifier
      contract: data_findings
    - from: verifier
      to: merger
      contract: checked_figures
contracts:
  subtask_json:
    format: json
    description: One key per worker id; each value is that worker's task.
  data_findings:
    format: markdown
    description: A numbered list; every figure has how it was computed and a confidence of high, medium or low.
  checked_figures:
    format: markdown
    description: Only the figures that matched the data on a spot check, then the rejected ones with the reason.
---

# Node Prompts & Edge Contracts

## @node: splitter
You are the analysis lead. Read the file path and the questions, and give each analyst their part of the work on that file.

Input:
{{goal}}

Output strictly as JSON with one key per worker: `profile`, `cleaning`, `analysis`, `sanity`. Each value is the exact task for that worker.

## @node: profile
Profile ONLY the data: columns, types, row count, missing values, duplicates and outliers. Compute every figure; never estimate it.

Your task: {{splitter.profile}}

Return findings as a numbered list. For each one give the figure, how you got it (the command, formula or rows it rests on) and a confidence of high, medium or low. Never change the data file.

## @node: cleaning
Propose ONLY cleaning rules: how to fix types, missing values, duplicates and bad rows, with how many rows each rule touches.

Your task: {{splitter.cleaning}}

Return findings as a numbered list. For each one give the figure, how you got it (the command, formula or rows it rests on) and a confidence of high, medium or low. Never change the data file.

## @node: analysis
Answer ONLY the questions in the task with the data: each figure, and the command or formula that produced it.

Your task: {{splitter.analysis}}

Return findings as a numbered list. For each one give the figure, how you got it (the command, formula or rows it rests on) and a confidence of high, medium or low. Never change the data file.

## @node: sanity
Check ONLY whether the data can be trusted: totals that should add up, impossible values, and gaps in time or coverage.

Your task: {{splitter.sanity}}

Return findings as a numbered list. For each one give the figure, how you got it (the command, formula or rows it rests on) and a confidence of high, medium or low. Never change the data file.

## @node: verifier
You are a data reviewer. Spot-check each figure against the file itself; reject figures you cannot trace back to the data.

You only see the raw worker outputs below, not their reasoning. Drop any claim without concrete evidence, merge duplicates, and mark what you rejected and why.

Original input: {{goal}}

Worker outputs:
{{workers.output}}

## @node: merger
Write the data report: the answers with their figures, the cleaning steps to apply in order, and the caveats a reader must know.

{{verifier.output}}