Convert a chat prompt
“Convert this classification or finite-label extraction prompt into EvalClimb decision questions and rules. Preserve my inputs and output labels, and compare with the original prompt's GPT-6 Luna baseline.”
Everything is public. No sign-in or API key needed.
Install the plugin, then start a new Claude Code session.
claude plugin marketplace add shpitdev/skills
claude plugin install evalclimb@shpitdev-skillsVerify: claude plugin list
claude mcp add --transport http evalclimb https://www.evalclimb.com/mcpVerify: claude mcp list
In Claude's connector settings, add a custom connector named EvalClimb with this URL and choose no sign-in.
https://www.evalclimb.com/mcpVerify: Enable the connector in a chat and ask to browse the public gallery.
Install the plugin, then start a new Codex session.
codex plugin marketplace add shpitdev/skills
codex plugin add evalclimb@shpitdev-skillsVerify: codex plugin list
codex mcp add evalclimb --url https://www.evalclimb.com/mcpVerify: codex mcp list
If your account or workspace allows custom MCP apps, enable developer mode and create an app with this URL. Choose no authentication, scan the tools and create the app.
https://www.evalclimb.com/mcpVerify: Enable the app in a chat and ask to browse the public gallery.
Merge this MCP entry into opencode.json, keeping any existing settings.
{
"mcp": {
"evalclimb": {
"type": "remote",
"url": "https://www.evalclimb.com/mcp",
"enabled": true
}
}
}Verify: Ask OpenCode to browse the EvalClimb public gallery.
Install the package in Pi 1.0+, then start a new session.
pi install https://github.com/shpitdev/skillsVerify: pi list
pi mcp add evalclimb --url https://www.evalclimb.com/mcp --exposure directVerify: pi mcp list
“Convert this classification or finite-label extraction prompt into EvalClimb decision questions and rules. Preserve my inputs and output labels, and compare with the original prompt's GPT-6 Luna baseline.”
“Tune this Jev request or Gemini/OpenAI decision spec on my labeled examples with EvalClimb. Keep the model and output contract, and show held-out gains, regressions and the exported caller.”
“Find EvalClimb's Gemini PDF invoice example. Walk me through the source PDFs, original and optimized instructions, and held-out mistakes. Help me adapt its native document spec to my task.”
“Compare compatible Jev, Gemini and OpenAI decision models in separate EvalClimb runs with the same native task, labeled examples, fixed learn/pick/test splits and optimization effort. Report untuned/tuned held-out scores and available cost/latency; never tune on test labels.”