hypertopos/hypertopos-skills

> Business data as geometry that AI agents navigate — the investigation workflows.
Install the skills into your agent, connect hypertopos-mcp to a sphere, and ask in plain language — the matching skill decides which tool to call, in which order, what the result means, and what to check next.
Judgement, not code. Nine Markdown skills — decision frameworks with worked examples — that an agent reads before it acts. No Python, nothing to run.
Every investigation ends in a report, not a tool dump. The skills carry a quality checklist, a report structure and anti-patterns, so a finding states its evidence, its confidence and its false-positive risk.
They know the tools' edges. Compact responses and the omitted map, sessions pinned to a snapshot, branches for what-if recalibration, and the parameter names a call must use.
Get started (one minute). Needs an agent that reads skills and a connected hypertopos MCP server:
bash
npx skills add hypertopos/hypertopos-skills # Claude Code, Cursor, Codex, Copilot, Amp, Cline, Windsurf, …
# or: copy the gds-/ folders into your agent's skills directory, e.g. .claude/skills/ for Claude Code
In Claude Code, the repository is also a plugin marketplace, so the skills install and update as a plugin:
/plugin marketplace add hypertopos/hypertopos-skills
/plugin install hypertopos-skills@hypertopos-skills
Then, in your AI assistant:
explore this sphere
why is account 85 anomalous?
check sphere health
run an AML scan
That's it. You get nine skills, each a folder with a SKILL.md (and references/ where the skill carries worked examples and detail the SKILL.md links to):
gds-analyst/ drives an investigation: goal, plan, delegation, report
gds-explorer/ orientation in a sphere you have never seen
gds-investigator/ root cause, entity deep-dive, hypothesis testing, false-positive assessment
gds-detective/ detection recipes: event rates, bursts, Simpson's paradox, collective drift
gds-scanner/ cross-pattern, neighbour-contamination, trajectory-shape and segment-shift scans
gds-monitor/ health checks, drift, regime changes, calibration epochs, safe recalibration
gds-fraud-investigator/ AML and fraud: screening, per-suspect investigation, typology recipes
gds-sphere-designer/ from a description of the data to a sphere.yaml, a build, and tuning
gds-spark/ anomaly detection on a Spark DataFrame: roles, find_anomalies, triage, explain, score
Works with Claude Code, Cursor, Codex, GitHub Copilot, Amp, Cline, Windsurf, and other supported agents.
> [!NOTE]
> The skills need hypertopos-mcp connected to a built sphere; they call its tools and read its responses. hypertopos builds the sphere. gds-sphere-designer works before a sphere exists, and gds-spark works in a Spark or Databricks notebook with hypertopos[spark], needing the MCP server only to hand a result to an agent.
## See it in action
A skill is text the agent follows. This is what gds-monitor says about recalibrating when the entities under review may be distorting the baseline (verbatim from its "Safe recalibration via branch" section):
text
# 1. Recalibrate on a branch (optionally excluding known influencers)
recalibrate(pattern_id, branch="my_branch", exclude_keys=["KEY1", "KEY2"])