eval-guide
Skill95 starsChanged 16 days agoMITView on GitHub
Entry point for eval work. Use when someone asks for help with evals in general, does not know where to start, asks "what should I do next", or asks for something no other skill in this plugin matches. Do NOT use when a more specific skill in this plugin already matches the request. Load that skill directly.
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 46
When in doubt about which row fits, ask one question instead of guessing.
≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Eval guide
This plugin splits eval work into skills, one per step. This skill only routes. Find the row that matches the user's situation, say which skill you are loading and why, then load it and follow its workflow. Do not improvise your own version of a step.
| The user's situation | Skill |
| Has a product, PRD or prompt, and no written definition of good | eval-rubric (v1) |
| Has a rubric and needs test inputs, or wants to know what a dataset is missing | uig |
| Has traces to review but no good way to read them, or the team isn't reviewing | trace-review-app |
| Has read traces and written notes | trace-codes |
| Has trace codes and a v1 rubric | eval-rubric (v2) |
| Has a rubric line or trace code that a rule can check | eval-code |
| Has a rubric line or trace code that needs judgment | eval-llm-judge |
| Has an LLM judge and human labels, and wants to know if the judge can be trusted | llm-align |
| Has a support ticket, complaint or bad trace to turn into test data | ticket-to-eval |
The order, for someone starting from nothing
eval-rubric: write v1 from the PRD.uig: build a diverse set of inputs and run the agent on them.- The user reads the traces and writes notes. If the traces are hard to read,
trace-review-appbuilds a review page first. trace-codes: cluster the notes into codes, theneval-rubricagain for v2.eval-codeandeval-llm-judge: one scorer per rubric line worth automating.- The user labels a sample pass or fail.
llm-align: measure each judge against those labels and fix it.ticket-to-eval: keep the dataset growing from real feedback.
Two steps stay with a person
Reading and annotating traces (step 3) and labeling for calibration (step 6) are done by the user. Do not do them on the user's behalf, and do not treat your own judgment as a label. If the user asks to skip either one, explain what the later steps lose: codes that reflect the model's guess about quality, and judges nobody has checked.
When in doubt about which row fits, ask one question instead of guessing.
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- microsoft/eval-guide★ 130same name
- kumaran-is/claude-code-onboarding★ 35same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- llms.txtllms.txt
- eval-codeSkill
- eval-llm-judgeSkill
- eval-rubricSkill
- llm-alignSkill
- ticket-to-evalSkill
- trace-codesSkill
- trace-review-appSkill
- uigSkill
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.