Skip to content
Game Market Copilot Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

gmc-analysis Skill

The gmc-analysis Skill gives coding agents guided workflows, statistical-validity rules, and chart-card guidance for analyzing Steam market data with Game Market Copilot.

gmc-analysis is a Skill for coding agents (Claude Code, Codex, and other agents that support a Skill or AGENTS.md-style instruction file). It packages the guided workflows, statistical-validity checklist, claim-safety rules, and chart-card conventions that make GMC output trustworthy to publish or act on.

The division of labor: the server owns data and access gates, the Skill’s recipes own mechanical collection and bounded aggregation, and the agent owns semantic interpretation, meaning it clusters theme labels, draws conclusions, and writes up findings. GMC does not cluster meanings across games for you.

Installing the Skill

Claude Code

npx skills add witchpot-studio/gmc-agent-skills

Via the gmc CLI

gmc skill install          # installs the Skill for Claude Code
gmc skill print --format agents-md >> AGENTS.md   # appends the Skill content for Codex

The Skill works against either the gmc MCP server or the gmc CLI. See /cli for CLI installation and authentication. Both paths carry the same workflows and honesty rules; only the tool names differ (for example, market_aggregate over MCP corresponds to gmc games aggregate on the CLI).

Authenticated access (CLI login, API keys, or the MCP server) requires a paid GMC plan. Anonymous CLI use is possible but returns redacted previews under a small quota.

Recipe families

The Skill organizes analysis into five recipe families:

  • R1, cohort complaints and praises. Size a cohort, gather per-title review evidence, and cluster recurring themes with validity guardrails attached.
  • R2, full-cohort map-reduce. For cohorts larger than one page of results: paginate, apply bounded concurrency, project only the fields needed, and reduce across chunks (including subagent chunking for large fan-outs).
  • R3, single-title deep dive. Profile one title, pull named review claims, check showcase fit, and position it against a comparable market segment.
  • R4, market sizing and distribution. Answer “how big” and “how distributed” questions using population-true server aggregates with explicit denominators, rather than estimating from a sample.
  • R5, repeatable cohorts via a saved game list. Persist a cohort definition (a filter or an explicit set of titles) once as a Game List, then reference it by ID across multiple calls instead of resending a large inline filter each time.

What the Skill is the source of truth for

  • Statistical validity: denominators and sample sizes on every share, naming the observation window or as-of date, distinguishing small-N groups as directional, and disclosing when a reported extreme came from scanning many buckets.
  • Claim safety: never using causal language for observational associations, always labeling figures not derived from GMC data as external estimates, and never treating not_collected as zero or as absent sentiment.
  • Chart guidance: when a result is turned into a shareable chart card, the required attribution, source labeling, and denominator/observation-window captions.

These rules are the same ones embedded in the guided MCP prompts (cohort_complaints, market_sizing, title_deep_dive, design_pillars), so an agent gets equivalent guardrails whether it uses the Skill directly or only has MCP access. See /methodology for the underlying rules in full.

When a recipe does not fit

If a documented recipe fails, or an agent discovers a better approach, the Skill asks the agent to record a short learning entry (what was tried, what happened, and a proposed rule) and submit it through the GMC feedback channel so the recipe can improve. This keeps the Skill’s guidance grounded in observed behavior rather than only in its original design.

  • /agents/mcp: the gmc MCP server and the read-only Docs MCP
  • /cli: CLI installation, authentication, and command reference
  • /methodology: coverage, sampling, and claim-safety rules
  • /troubleshooting: common errors and what they mean

Was this page helpful?