Skip to content

nanocoai/analyst

v1.0.0

Data analyst agent: pipeline checks, query writing, and recurring report integrity

GitHub

Data Analyst Template

A NanoClaw template for a data analyst / reporting assistant serving one principal: keep the reporting pipeline healthy, keep metric definitions consistent, and turn report requests into buildable, validated deliverables. It ships as a seed and grows around one principal's stack — no bundled tools, no hard-wired databases.

Layout

analyst/
├── plugin.json                        # Agent Plugins manifest (marks the folder as a plugin)
├── ai.nanoco.nanoclaw/
│   ├── context/
│   │   ├── instructions.md            # the persona — proactive, fact-first, close the loop
│   │   └── additional_context/
│   │       └── memory-structure.md    # where learned state lives: metrics, pipelines, audiences, systems
│   └── tasks/
│       ├── morning-pipeline-check.md  # weekdays at 07:00 — pipeline status before anyone opens a report
│       └── weekly-report-integrity.md # Mondays at 09:00 — metric drift and report disagreements
├── skills/
│   ├── pipeline-check/
│   │   └── SKILL.md    # verify the scheduled data work ran and produced sensible output
│   ├── query-writing/
│   │   └── SKILL.md    # SQL/MongoDB queries: right grain, no fan-out, checked against knowns
│   ├── report-onboarding/
│   │   └── SKILL.md    # stand up reporting for a new audience (client, BU, region) end to end
│   ├── report-spec/
│   │   └── SKILL.md    # turn a report/widget request into something buildable
│   └── schema-and-cleanup/
│       └── SKILL.md    # fix bad data and change shape without breaking readers
└── README.md                          # this file

Memory

Learned state does not live in the template — it lives in memory/, described by ai.nanoco.nanoclaw/context/additional_context/memory-structure.md and built as the agent works. The skills and tasks read from and write to these paths:

  • memory/principal.md — who the analyst serves, the platforms and pipelines they own, and where the line sits between handled and escalated.
  • memory/conventions/metrics.md — one entry per metric: definition, source, and where it is computed. Metrics computed in two places are the root of most reporting disagreements, so this file records where each one lives.
  • memory/conventions/pipelines.md — every scheduled job with its normal output volume per audience; this is what the morning check measures against.
  • memory/conventions/audiences.md — one entry per reporting audience (an external client, a business unit, a region, an exec team): their reports, scoping rules, and what was validated against numbers on their side.
  • memory/conventions/systems.md — schemas, ambiguous-field meanings, and what reads each table from outside the application.
  • memory/queries/ — saved queries with the question each answers; a one-off pull is never a one-off.
  • sources/ — the immutable raw record (schema exports, query outputs, dumps).

Stamp an agent

ncl groups create --template data/analyst --name "Data Analyst"

Wire the agent to a channel (/manage-channels) and connect it to the stack it should work against.

Notes

  • No mcp.json. Every analyst's stack is different (Postgres, MongoDB, a BI tool, a warehouse). Add servers as needed via the /add-*-tool skills; OneCLI injects credentials at request time, so no secrets live here.
  • Fact-first by default. The agent reads the live table, thread, or file before answering, carries exact values through unchanged, and flags anything unverified.
  • Scheduled routines start paused. The weekday pipeline check and Monday integrity review are installed with the template and run only after the user activates them. Both report and draft fixes; neither patches data by hand.
  • It grows itself. When the agent repeats a procedure, it writes a new skill under skills/<name>/, and durable conventions go in memory/conventions/.