Claude Skills by DaveGold

Guides building a NEW rich-domain MCP server or tool, and auditing an EXISTING one up to a measured reference, using Introspective Context Engineering: discover the domain from live data, name fields so they cannot be misread, deliver guidance where the model actually receives it (description head, input schema, response — not the output schema, not past char 2,048), compute determinate verdicts, record provenance per rule, and measure the result. Invoke when asked to add, build, scaffold or ...

Run the eval set in evals/questions.json against the arm servers (thin/words/rich, or opaque/opaque-words) and score the results. Use when asked to run the eval, score an arm, compare arms or models on the building-profile eval, or reproduce the numbers for the talk.

Checks in about a minute whether an MCP server's metadata survives the clients it will run in (Claude Code, claude.ai, Cowork, ChatGPT Chat and Work, Codex): reads tools/list and the server instructions over HTTP or stdio, and scores each tool against each client's measured limits on descriptions, server instructions, input schemas and output schemas. Invoke when asked to check, health-check, lint or score an MCP server, to ask "will this server work in ChatGPT / Codex / claude.ai", "why does...

Guides building a NEW rich-domain MCP server or tool, and auditing an EXISTING one up to a measured reference, using Introspective Context Engineering: discover the domain from live data, name fields so they cannot be misread, deliver guidance where the model actually receives it (description head, input schema, response — not the output schema, not past char 2,048), compute determinate verdicts, record provenance per rule, and measure the result. Invoke when asked to add, build, scaffold or ...

Run the eval set in evals/questions.json against the arm servers (thin/words/rich, or opaque/opaque-words) and score the results. Use when asked to run the eval, score an arm, compare arms or models on the building-profile eval, or reproduce the numbers for the talk.

Checks in about a minute whether an MCP server's metadata survives the clients it will run in (Claude Code, claude.ai, Cowork, ChatGPT Chat and Work, Codex): reads tools/list and the server instructions over HTTP or stdio, and scores each tool against each client's measured limits on descriptions, server instructions, input schemas and output schemas. Invoke when asked to check, health-check, lint or score an MCP server, to ask "will this server work in ChatGPT / Codex / claude.ai", "why does...

Guides building a NEW rich-domain MCP server or tool, and auditing an EXISTING one up to a measured reference, using Introspective Context Engineering: discover the domain from live data, name fields so they cannot be misread, deliver guidance where the model actually receives it (description head, input schema, response — not the output schema, not past char 2,048), compute determinate verdicts, record provenance per rule, and measure the result. Invoke when asked to add, build, scaffold or ...

The skill is meant to be copied into other repos **unchanged**, so that a later sync is a copy, not a merge. Your own knowledge (platform and deploy wiring, auth, vendor quirks, conventions, telemetry, your own reference implementations, your own evidence) goes in overlays that the skill reads at fixed points. 1. **Copy the skill folder** (`.claude/skills/rich-domain-mcp-server/`, and the `.codex/` copy if you use Codex) and record the commit or tag you took. Pin a tag (`skill-v1.1.0`, …) rat...