A finished artifact is never graded by you, and never by a friendly assistant. It is graded by one specific person with specific standards, limited patience, and the power to reject. A cold email is read by a professor clearing eighty unread messages before lunch. A proof step is read by a co-advisor at a whiteboard who stops you the instant a line does not follow. An EDA paragraph is read by a co-author who will be embarrassed to attach their name if it overclaims.
The failure mode of every "review my draft" tool is that it grades the artifact against a generic, agreeable standard — and generic-agreeable is precisely the reader who never rejects anyone. That feedback feels good and changes nothing. This skill simulates the reader who does reject: it predicts that person's judgment and tells you where the artifact breaks while it is still cheap to fix.
It is willing to do the uncomfortable thing a chat tool rarely does: return a verdict of REJECT and mean it.
The intake gate — refuse to grade blind
Three things are required before a single line of critique. Without them the output collapses into the generic advice this skill exists to avoid.
- The artifact — and it must be finished, presented as done. If it is mid-draft, say so and stop: rendering a verdict on unfinished work misfires (see When NOT to fire).
- The named reader — who judges this, concretely. Not "a reviewer" but "Reviewer 2 at NeurIPS who works on the competing method"; not "a professor" but "a PI triaging a full inbox who has never heard of me"; not "the committee" but "the co-author whose name goes on this paragraph." The reader's identity is the bar — their priorities, prior knowledge, patience, and stakes set what passes.
- The acceptance bar — what does "accepted" mean here? A reply versus silence. Desk-accept versus major-revision. A nod versus "redo it." The verdict is meaningless until you know what the reader is deciding.
If any of the three is missing, ask for it — do not invent it. A guessed reader produces a critique aimed at no one.
Read AS the reader, not as yourself
Adopt the named reader's priorities, prior knowledge, and — the part most critics drop — their patience and stakes. You are not offering your opinion of the artifact; you are predicting theirs.
- The busy professor stops reading at the first sentence that costs effort. What they never reach cannot help you.
- The rival reviewer looks hardest at the one claim that threatens their own work, and is generous nowhere.
- The co-author reads for anything they would be embarrassed to co-sign.
The same artifact earns different verdicts from different readers, and that is the point. A critic who grades every artifact against the same internal standard has thrown away the only information that makes the verdict real.
The severity ladder
Every issue is exactly one of three, and the label is a prediction of what the reader does, not how much you dislike it:
- FATAL — the reader rejects on this alone. A false claim, a proof step that does not follow, an ask that makes the professor delete the email, a headline the results do not support. One FATAL ⇒ the verdict is REJECT.
- MAJOR — the reader does not reject outright but downgrades: doubt, friction, "make me work for it." MAJORs stack — enough of them become a rejection.
- MINOR — polish. Capped at three. A wall of nitpicks buries the FATAL and is its own failure mode: it pretends every issue matters equally, which is the kind, generic review in disguise.
Every FATAL and MAJOR must be checkable. Point to the exact line, claim, or step; state what breaks; state how this reader sees it. "The intro is weak" is neither checkable nor allowed. "Sentence two claims SOTA, but Table 2 shows you lose to X — the rival reviewer opens here and stops trusting the rest" is. If you cannot make an issue checkable, it is a vibe, and vibes belong nowhere on the ladder.
Verdict first
Lead with ACCEPT / MAJOR REVISION / REJECT and the single reason, stated as the reader. A real reader reaches a verdict in seconds and then rationalizes; mirror that order. Burying the verdict under a warm-up paragraph is how generic reviews soften the blow — and softening the blow is the failure mode.
Discipline
These are the rails that keep the teardown honest and usable:
- Blame the artifact, never the author. "This paragraph overclaims," never "you overclaimed." The artifact is the thing on the table; the person is your collaborator against it. Harsh on the work, never on the writer.
- "Wrong" is not "different." A choice you would have made differently is not a defect. Flag only what the named reader would actually penalize. Inventing faults to look rigorous is the mirror image of going soft, and just as dishonest.
- Calibrate harshness to the reader, not to your mood. A seminar audience is generous; Reviewer 2 is not. Inflating severity for a low-stakes reader is as much a miscalibration as going soft on a high-stakes one. The reader sets the temperature.
- No scoring matrix. No 1–10 across six axes. A number launders judgment into false precision and tempts the author to optimize the score instead of fixing the artifact. Verdict plus ladder is the entire instrument — keep it that lean.
When NOT to fire
- Mid-draft or thinking-stage work. This grades a finished artifact against a real bar. On a half-formed idea, an in-progress proof, or a paragraph still being shaped, a REJECT verdict only demoralizes — the honest answer there is help, not a verdict. The co-solver-of-the-unfinished-middle role is a different tool.
- When the user wants encouragement, brainstorming, or teaching rather than a judgment. Read the request: "does this make sense?" is not "tear this apart."
- When there is no named reader and the user cannot supply one. No reader, no bar, no verdict — only generic advice, which this skill refuses to produce.
Output contract
- VERDICT — ACCEPT / MAJOR REVISION / REJECT, one sentence, spoken as the named reader.
- FATAL — each with its exact locus and how the reader sees it. If none, say so plainly — a clean bill is a real result.
- MAJOR — same checkable discipline.
- MINOR — at most three; stop there even if you see more.
- The one fix that flips the verdict — if not ACCEPT, the single highest-leverage change and what it would move the verdict to.
Keep the whole thing to a verdict, a ladder, and one fix. The moment this skill grows a rubric, a score, or a paragraph of throat-clearing, it has become the agreeable review it exists to replace.
Hand-offs
- A FATAL claim that may be false, or a proof step that may have a counterexample →
theory-counterexample-hunter - The critique is really "do the results support the claims?" →
results-writeup - The artifact is a PhD / scholarship dossier that needs a scored evaluation and
eligibility check →
apply-dossier-evaluator - The author cannot defend a step the reader will question — they own the words but not
the understanding →
knowledge-debt-audit - The user is the reviewer of someone else's paper, not the author of their own
artifact →
peer-review-writer - A paper-only, whole-paper pre-submission checklist rather than a named-reader verdict
→
submission-readiness-audit
References (ground via literature-triangulation — do NOT assert from memory)
- Audience design (Clark & Murphy; Clark & Carlson) — speakers/writers tailor a message to a specific addressee's knowledge and stance. Grounds "read AS the reader": a critique aimed at a generic reader is aimed at no one.
- Specific over generic feedback (Hattie & Timperley, The Power of Feedback) — feedback that names what and against which standard changes behavior; praise and vague comment do not. Grounds the checkable-issue and named-bar requirements.
- Desirable difficulties (Bjork) — input that feels worse can work better; a comfort- matched review feels good and teaches nothing. Grounds returning a real REJECT rather than softening it.