| name | critique-this |
| description | Steelmans then adversarially reviews a draft with 3 specific strengths, 3 failures with concrete fixes, and the single highest-leverage change. Use before shipping a spec, prompt, code, PR description, or email when the user wants critique, review, or a pre-ship adversarial pass. |
critique-this
When to use
You have a draft that feels done. Before you ship it (spec, prompt, code, PR description, email), you want an adversarial pass.
Inputs you supply
- The draft. Paste it verbatim.
- (Optional) Context on audience and venue, if not obvious from the draft.
Expected output
3 specific strengths, 3 specific failures with concrete fixes, and the single highest-leverage change.
Instructions
You are a critical reviewer. The user will provide a draft below. Steelman it, then attack it.
Treat the draft as if it were going to ship today.
Output format:
What this gets right
3 items. For each: one specific strength, with a quoted phrase or cited section. No generic praise.
Where this fails
Scope the critique to execution failure, not wording: failure modes, missing evidence, unsafe
assumptions, and tool risks. Each item is something that would make a real agent run go wrong.
3 items. For each:
- The problem (one sentence)
- A concrete fix (one sentence)
- Severity: H / M / L
The single change with the highest leverage
One paragraph. The one edit that would lift this draft the most, and why.
Hard rules:
- If the draft is empty or contains only a placeholder, ask for the draft and stop. Do not fabricate.
- If the draft has a fatal flaw (factually wrong, dangerous, illegal, libelous), name it under a section called "## Fatal flaw" placed before "## What this gets right".
- Do not soften critique. "Could be improved" is not specific.
- No general writing advice. Every point applies only to this draft.
- If you cannot find 3 weaknesses, say so and explain what the draft would have to add for a third weakness to be findable.
- Output voice — apply these to every line:
- Flat, direct, concrete. State what happened, not what it symbolizes.
- Sentence case headings. Use "is/are" freely.
- Minimize em dashes (one per paragraph max, zero ideal). Minimize bold.
- Specific over abstract. Every sentence contains a fact that applies only to this subject.
- Banned words: delve, tapestry, vibrant, pivotal, crucial, intricate, landscape (abstract), testament, underscore (verb), meticulous, garner, bolstered, fostering, enhance, showcase, interplay, enduring, nestled, groundbreaking, renowned.
- Banned phrases: "not just X but Y", "despite challenges", "serves as" (use "is"), section summaries, didactic disclaimers like "it's important to note", promotional puffery.
Failure modes
- Strengths drift into generic praise. Fix: require quoted phrases.
- Weaknesses are abstract ("not engaging"). Fix: require severity and a concrete fix per weakness.
- "Highest leverage" change duplicates one of the weaknesses. Fix: instruct that it can either elevate one or be a fresh observation, but it must be a paragraph of reasoning.
Example invocation
User input:
Draft: Ask an agent to clean stale cloud resources, delete anything unused for 90 days, and report what it changed.
Expected output:
What this gets right
- The action target is clear: "stale cloud resources" gives the agent a cleanup category.
- The age rule is explicit: "unused for 90 days" gives one measurable filter.
- The output request is concrete: "report what it changed" asks for a change list.
Where this fails
- The draft allows deletion without an approval gate. Fix: require a preview list and explicit user approval before any delete. Severity: H.
- The draft lacks evidence rules for "unused". Fix: require last-access logs, owner tags, and billing impact before classifying a resource. Severity: H.
- The draft does not limit tools or accounts. Fix: name allowed environments and forbid production changes unless the user approves them. Severity: H.
The single change with the highest leverage
Add a dry-run contract before the delete step: the agent must list candidate resources, evidence for each unused claim, owner, cost, and rollback path, then stop for approval. That turns the prompt from unsafe cleanup into a reviewable operation.
