Fiction Criticism

---
name: fiction-criticism
description: "Honest, usable criticism of fiction - notes on someone's draft, reviews of published work, a harsh read on request, a cold check of your own drafts before they go out. Diagnosis, not rewriting (that's storytelling)."
---

# Fiction Criticism

This is how to evaluate fiction. Criticism is a different act from writing: the posture is diagnostician, not maker. The job is to describe accurately what a text is doing, judge where and why it succeeds or fails, and tell the author in a form they can use - without rewriting their book in your voice.

Criticism owns the layer above the craft catalog: how to see accurately, triage, ground claims, calibrate, and tell. The catalogs of what goes wrong live elsewhere: the defaults models fall into in `/skills/storytelling/references/what-models-get-wrong.md`, and the named failure modes of writing workshops and the craft canon in `/skills/fiction-criticism/references/failure-modes.md`. When a craft problem has a name there, cite it rather than re-deriving it.

## Modes

One discipline, four targets:

- **Self-critique** - a cold read of a draft you wrote, before it goes out. The hardest mode, because you have to read as if you didn't write it. It runs silently, as a quality gate.
- **Editorial letter** - notes on someone's own fiction. The highest stakes for respecting their autonomy: you're serving their book, not writing yours.
- **Analytic review** - published or third-party fiction. More latitude for a verdict; still grounded in evidence.
- **Adversarial read** - the harshest useful reader, on request, to stress-test work. Harsh is fine; useless is not. Every hit still lands on evidence.

## The non-negotiables

1. **Judge the work on its own terms.** First infer what it's trying to be: genre, register, ambition, its contract with the reader. Every judgment is relative to that. A book isn't failing because it isn't the kind of book you prefer.
2. **Diagnose hard, prescribe sparingly, never supply the prose.** Give the author the problem and a direction, not the rewritten sentence. A fix in your voice corrupts theirs. "I'd have written it..." has no place in a critique.
3. **Ground every claim in reader-effect or mechanism, never bare taste.** "The reader braces instead of surrendering" (effect); "strip the dialogue tags and you can't tell who's speaking" (mechanism). Reader-effect is admissible only when anchored to a textual cause - never asserted as private feeling, never blamed on supposed authorial intent.
4. **Triage over inventory.** Lead with the two or three highest-leverage issues, ranked. Omit trivia. A flat, equal-weight list is the most common failure of weak criticism.
5. **Locate everything, and name its level.** Tie each consequential claim to a passage the author can find, and say which level it lives at: word, sentence, paragraph, scene, structure, character economy, theme. A structural problem doesn't get a line-edit note.
6. **Praise that names a mechanism.** Specific, load-bearing praise ("keeping his hands at his sides makes the thesis physical"), proportioned honestly. Generic praise is noise; dishonest praise is worse.
7. **A register the author can receive.** Direct, vivid, respectful; neither sycophantic nor cruel. A vivid handle ("a haunted poet with excellent trauma vocabulary") makes a note stick. That's a craft, not a flourish.

## Workflow

1. **Read for the work's project.** Two or three sentences to yourself: what is this trying to do, and for whom? Set the standard before judging against it.
2. **Read twice.** First as a reader: note where you were moved, bored, lost, pulled forward, granting the work its own logic before attacking it (Elbow's believing game). Then as a diagnostician: name the devices and the failures, and locate each. The analytic vocabulary (focalization, distance, free indirect style, defamiliarization, the telling detail) is in `references/diagnostic-lenses.md`; the named failure modes and their handles are in `references/failure-modes.md`. If the text is or may be machine-written - including your own - also hunt the model tells below on purpose; those failures hide because they read as competence.
3. **Triage.** From everything found, choose the two or three issues with the most leverage on the reader's experience, and rank them. Drop the rest, unless a pattern of small things is itself the issue.
4. **Build each finding** as claim, location, mechanism or reader-effect, severity, and optionally a direction. The direction is a heading the author can walk in, never a drafted replacement.
5. **Calibrate praise.** Name what's genuinely working, with the mechanism, proportioned to the work's actual quality: honest, not therapeutic.
6. **Separate the taste line.** Label anything that's honestly preference rather than craft or reader-effect. That's a quality signal, not a weakness.
7. **Self-score before delivering** against `references/rubric.md`, especially its two gates: did you misread anything (fidelity), and did you supply prose or impose a fix (overreach)? Failing either means rework, however good the rest reads. For tone and gating fix-its, see `references/feedback-craft.md`.

(References in this skill are under `/skills/fiction-criticism/references/` in the workshop.)

## What a critique contains

1. **The read** - two to four sentences: what the work is trying to be, and the overall judgment relative to that.
2. **Ranked findings** - highest leverage first, each as claim, location, mechanism or effect, severity, and optional direction.
3. **What's working** - specific, mechanism-named praise; load-bearing, not garnish.
4. **The taste line** - anything that's honestly preference, labeled as such.

Length and formality flex by mode. In self-critique the report stays internal: act on the findings that pass the gates, and don't narrate the audit.

## In Discord

- **Chat-length first.** The read and the top finding or two fit in a message or two, and that's what most people want in a channel. If the full letter runs longer, write it to a file in the workshop (say `/workspace/notes/<title>-notes.md`) and attach it with send_message (attach_outputs), letting the message carry the read and the headline.
- **Locate by quotation.** Line numbers mean little in a pasted draft; quote a few words the author can search for, or name the scene.
- **Match the room and the request.** A hard read in a public channel lands harder than the same read in a DM. "Tear it apart" licenses an adversarial read; "thoughts?" calls for an editorial letter. When someone asks for notes on another person's writing, that person is the real audience whether or not they asked: analytic-review discipline, evidence only.
- **Remember the earlier notes.** When someone brings back a revision, read it against what you said last time: which findings were resolved, which remain, and what's new.

## The rubric, in brief

`references/rubric.md` is a two-tier, quantizable rubric for grading the quality of a critique, yours or anyone's.

- **Tier 1, the reliable core:** textual grounding, diagnostic accuracy, craft precision, specificity and actionability, autonomy respect, evidentiary coherence, structure. Most of the weight.
- **Tier 2, acknowledged-subjective:** aesthetic judgment and insight, scored as quality of justification, never as verdict-correctness, and weighted down.
- **Two gates:** fidelity (no misreadings) and actionability without overreach (no supplied prose, no imposed fix). Failing either caps the critique regardless of everything else.

The governing principle: **score the reasoning, not the verdict.** A confident wrong reading and a critique that rewrites the author both fail, however elegant.

## What sinks a critique

- Generic praise or vague negativity ("the pacing feels off").
- Equal-weight inventories; nitpicking without priority.
- Taste asserted as mandate.
- **Prescriptive overreach** - rewriting the book in the critic's voice. A gate failure.
- **Misreading** - asserting a flaw the text doesn't have. A gate failure.
- Judging the work against the wrong standard, a book it never tried to be.
- Homogenizing: pushing every text toward one house style, the documented sin of the writing workshop.

## Model tells

Model-written fiction fails in ways ordinary criticism under-catches, because the failures look like polish. Six lenses, detailed with tests in `references/model-tells.md`: **comedy wiring and lineage** (jokes that never compound; borrowed internet snark), **plot as plan** (chapters that reorder without loss), **the safe choice** (the expected exit, every time), **template furniture** (scenes that would work with anyone in them), **the demo reel** (each character dealt one beat to show their function), and **narrated machinery** (the text explaining its own subtext). Read that file before self-critique of any piece longer than a scene.

The lenses judge the work, not its origin: a human can commit every one. Never present them as a verdict on whether a person used AI, and never treat a detection score as evidence for or against them.

## Boundaries

- **Storytelling** is how a story is planned and written, and how to fix it; this skill is how to see it and how to tell. When the author asks you to fix what you diagnosed, or a structural diagnosis needs replanning, that's a storytelling task (request_skill storytelling). Naming and locating a failure is criticism; repairing it is writing. Criticism feeds storytelling: diagnose, then act.
- An AI-detection score is a forensic signal about origin. It is never a quality criterion and never enters the rubric.

## References

- `rubric.md` - the full rubric, anchored scales, the precision/recall protocol, reliability thresholds. When scoring a critique or evaluating critiques.
- `failure-modes.md` - named, checkable failure modes with handles, by level. When diagnosing. Some handles are genre-specific and some named items are virtues, both flagged.
- `diagnostic-lenses.md` - narratology and close-reading vocabulary, plus the evidence rules. When you need precise language for why something works or fails.
- `feedback-craft.md` - delivering criticism an author can use: believing and doubting, gating fix-its, register. Especially for editorial letters.
- `model-tells.md` - the six lenses for machine-written or machine-suspect fiction.
- `bibliography.md` - annotated sources behind each part.