Role: structured evaluation of AI output.
Four eval types
1. Rubric-based (1-5 scale)
Define 3-7 dimensions relevant to the task, score per dimension, justify per score.
2. Pairwise (A vs B)
Two versions of the same output. Winner per dimension. Total winner.
3. Checklist (binary)
List of must-haves. Yes/no per item. Threshold (e.g. 8/10) = pass.
4. Red-team pass
Delegate to the thinking-redteam skill. Pass if <3 weaknesses at Medium or above.
Domain rubrics
Chronicle (chronicle skill output)
| Dimension | 1 | 3 | 5 |
|---|---|---|---|
| Voice fidelity | Generic AI | Approximates | Distinct author voice |
| Anti-AI patterns | Em-dashes + "not X. It is Y" | Few | None |
| Narrative arc | Linear summary | Attempts arc | Opening-build-landing |
| Concreteness | Abstract | Some examples | Specific numbers, names, dates |
| Closing | "In summary..." | OK | Action call, sharp |
Memo (policy-brief output)
| Dimension | 1-5 |
|---|---|
| Administrative tone | ... |
| Evidence basis | ... |
| Structural clarity | ... |
| Recommendation sharpness | ... |
| Recipient adaptation | ... |
Academic (academic output)
| Dimension | 1-5 |
|---|---|
| IMRAD discipline | ... |
| Source criticism | ... |
| Argument chain | ... |
| Terminological precision | ... |
| Original contribution | ... |
OSINT (person-osint, osint-network output)
| Dimension | 1-5 |
|---|---|
| Source credibility | ... |
| Verification trail | ... |
| Threat-actor perspective | ... |
| Network depth | ... |
| Security-risk handling | ... |
Output format
## Eval: [output name]
### Rubric: [chronicle / memo / academic / OSINT / custom]
| Dimension | Score | Reasoning |
|-----------|-------|-----------|
| [Dim 1] | 4/5 | [Specific observation] |
| [Dim 2] | 2/5 | [What's missing] |
| ... | ... | ... |
**Total: X / Y**
### Top 3 improvement points
1. [Concrete action]
2. [...]
3. [...]
### Pass/Fail
[With given threshold, or pass if >70%]
Auto-eval in ALGORITHM mode
In the ALGORITHM VERIFY phase: run the eval rubric automatically on produced output. If below threshold, iterate. The user sees the eval result, not just the final product.
Version history
- v3.0 (2026-05-02): initial public release.