| name | critic |
| description | Adversarial review of the most recent piece of work in the current session, delegated to a sub-agent with fresh context. Probes for errors, untested assumptions, and glossed-over decisions. Produces a short critical document for the user. Use when asked to "critique", "be skeptical", "challenge me", or "find what's wrong". |
Critic
A skeptical second pair of eyes on the work just produced -- find what is wrong, weak, or unjustified before the user does.
Procedure
-
Identify the artefact. The most recent piece of work in this session: a function, model, figure, paragraph, data-cleaning step, methodological choice. If multiple candidates are plausible, name them and ask.
-
Launch a sub-agent with fresh context. Hand it the artefact and the minimum context (what the user is doing, unit of analysis). No defence of the current approach.
-
Sub-agent produces a short critical document (250--500 words). Format below.
-
Save and report. Write to critiques/YYYY-MM-DD_HHMM_<short-slug>.md and report the path. Do not summarise or soften.
What to probe
- Untested assumptions.
- Stated but unverified facts.
- Logical / inferential gaps.
- Glossed-over decisions.
- Failure modes not considered.
- Sycophancy in the original work.
What NOT to do
- Do not validate. If nothing material is found, say so and list what was checked.
- Do not propose fixes. The critic finds problems.
- Do not be generic. Cite line numbers, variables, sentences.
- Do not pile on minutiae. Five real problems beat fifty stylistic nits.
Output format
# Critique: <one-line topic>
**Artefact:** <file path(s) or description>
**Date:** <YYYY-MM-DD HH:MM>
## Findings
1. **<short label>.** 2--3 sentences. Cite specifically.
2. ...
## Assumptions left unverified
- <one-line each>
## What was checked and found OK
- <Brief.>