breakthrough-evaluator-agent

Original：🇺🇸 English

Not Translated

Send an independent judge that is not the author: a fresh evaluator subagent that checks finished work against pass criteria written before it was built (evaluator mode), or audits a batch against its source of truth (audit mode), with evidence for every verdict and never a fix. Use when the owner asks for an evaluator, reviewer, auditor or judge, or says 'send a judge', 'evaluate this', 'check my work with a fresh agent', 'get a second agent to review this', 'audit this folder', 'audit my registry', 'verify this against the source of truth', or 'LLM-as-judge'. Runs only when asked; it does not fire on its own at the end of a task. Not for checks a machine can run (tests, schema validation, a search that must return zero), not for hundreds of identical units, and not for fixing anything.

7installs

Sourcebreakthrough-edu/breakthrough-evaluator-agent

Added on2026-10-24

NPX Install

npx skill4agent add breakthrough-edu/breakthrough-evaluator-agent breakthrough-evaluator-agent

SKILL.md Content

Evaluator agent

This skill is a pointer. ⛔ It holds no rules, no brief template and no output schema: all of that lives in the playbook, in one place, so there is nothing here to drift out of date. Do the three steps below.

1. Find the playbook, the owner's copy first

A my-second-brain vault. If the working directory (or a folder the owner names) has
```
99_Meta/structure-doctrine.md
```
, read that doctrine's section 1 for the real path of the methodology layer, then look for
```
<methodology>/Playbooks/evaluator-agent/
```
. If the folder is there, enter through its door: read
```
_Evaluator-Agent-Guide.md
```
and do its four beats, then read
```
evaluator-agent-playbook.md
```
in the same folder. The owner's copy always wins over the seed.
A copy the owner keeps elsewhere. Look for a standing pointer that names the folder of the owner's copy: in your memory, or in the
```
CLAUDE.md
```
of the folder you are working from (the installer writes one there when the owner saves a copy outside a my-second-brain vault). If you find one, or the owner tells you where the copy is, read the door and the playbook in that folder.
Otherwise, the seed shipped with this skill:
```
playbook/evaluator-agent/evaluator-agent-playbook.md
```
, relative to this file. Tell the owner in one line: "Using the seed playbook that came with the skill; nothing from this run will be kept unless you save a copy." The seed works without saving anything, but updating the skill replaces it, so the three jobs and the run log only last in a saved copy. When running from the seed, skip step 6 of the playbook's "The moves" (registering a row on the door) and ⛔ never edit any file inside this skill's folder, because an update replaces them.

If the playbook's "Your three real jobs" still shows placeholders, ask the owner for them once. ⛔ It does not block the run: if they skip it, carry on with the job in front of you.

2. Send the judge

Follow step 1 and step 2 of the playbook's "The moves": fill the six field brief, then send a new general-purpose subagent with the brief and nothing else. ⛔ Never a fork.

3. Receive the report

Follow the playbook's "The moves" from step 3 on. It starts with rerunning one piece of the judge's evidence yourself, before you trust anything else in the report.

Not installed properly, or want it inside a vault? The installer is

INSTALL.md

at https://github.com/breakthrough-edu/breakthrough-evaluator-agent.