story_diagnosis_loop v1.0.0

name: story_diagnosis_loop description: Falsification-driven story review for fiction manuscripts and scripts. Builds competing diagnoses of what's wrong, anchors each to quoted evidence, tries to kill each diagnosis, stress-tests it with counterfactuals, re-checks blind, then outputs revision task cards. Triggers when someone asks to critique a manuscript, review a screenplay scene, find the real problem in a draft, or turn vague reader feedback into fixable work.

Story Diagnosis Loop

Purpose

Vague feedback ("the middle felt slow", "I didn't connect") is a symptom, not a diagnosis. A story review that stops at symptoms prescribes the wrong fix. This skill runs a falsification loop on a manuscript or script: propose several competing diagnoses for the core problem, then try to disprove each one before accepting any of them. What survives the loop is the real problem, and each surviving diagnosis ships as a revision task card the writer can actually execute.

Works on full drafts, single scenes, short stories, screenplay pages, or a stack of beta-reader notes.

Workflow

Run the passes in order. Do not skip a pass because a diagnosis "feels right."

Pass 1: Intake

1. Read the material twice. First read: reader's read, no notes, mark only where your attention dropped or your body tensed. Second read: craft read, against the rubric (see references/story-craft-rubric.md).
2. If the writer supplied beta notes, log each note as a raw symptom. Never treat a symptom as a diagnosis.
3. Record the writer's stated intent (genre, audience, what the story is trying to do). This is the yardstick. You are reviewing this story on its own terms, not rewriting it into a different one.

Pass 2: Competing diagnoses

Draft 2 to 4 diagnoses, and they must genuinely compete (different root causes, different fixes). A good set:

• Diagnosis A: a structural problem (causality break, missing scene shift, stakes collapse).
• Diagnosis B: a character problem (agency, motivation, inconsistent behavior).
• Diagnosis C: a craft problem (exposition load, POV distance, dialogue on the nose).
• Optional Diagnosis D: the contrarian one (what if the "problem" is actually working and the real issue is elsewhere).

Write each diagnosis as: claim (one sentence), mechanism (how it produces the symptom), fix shape (what kind of revision would address it). No evidence yet. Vague claims get cut here: if you cannot state the mechanism, the diagnosis is a mood, not a diagnosis.

Pass 3: Evidence anchors

For each diagnosis, attach 2 or more anchors: short direct quotes from the manuscript (with scene/chapter and page or line markers) that the diagnosis predicts. Then attach the counter-anchors: quoted passages that should NOT exist if the diagnosis were true.

Rules:

• An anchor is quoted text from the manuscript, never paraphrase. Paraphrase is where reviewers smuggle in their own reading.
• Every diagnosis must have at least one counter-anchor. If you cannot find one, you are not looking hard enough; go find the passage that makes the diagnosis uncomfortable.

Pass 4: Falsification (the core of the loop)

Now try to kill each diagnosis:

1. Strongest counter-reading. For each diagnosis, write the best argument that it is wrong, using only the counter-anchors. If the counter-reading is stronger than the diagnosis, kill the diagnosis.
2. Scope test. Does the diagnosis explain ALL the symptoms, or just one? A diagnosis that explains one flagged scene but breaks when applied to the rest of the manuscript is overfitting. Kill or narrow it.
3. Minimum-viable version. Shrink each surviving diagnosis to its smallest defensible claim. "The middle sags because the protagonist has no scene-level goal for six chapters" beats "the pacing is off."
4. The reader-honesty check. Ask: is this my preference talking? If the fix would turn this story into a different story (one you would rather have read), the diagnosis fails. Review the story on the page, not the one in your head.

At the end of this pass, expect 1 or 2 survivors. Zero survivors is a valid outcome: it means the symptoms were misread. Go back to Pass 1.

Pass 5: Counterfactual tests

For each surviving diagnosis, run one test:

• Removal test: "If we deleted the flagged material entirely, would the story still work?" If yes, the material was dead weight, not broken.
• Inversion test: "If the protagonist got what they wanted here, would anything break downstream?" If nothing breaks, the stakes were decorative.
• Substitution test: "If a different character made this choice, would the scene still play?" If yes, the choice wasn't character-driven.
• Compression test: "Could this scene be cut in half without losing its shift?" If yes, say so precisely (what to cut, what must stay).

Each test must end with a quoted passage and a yes/no plus one sentence of why. A test you cannot run is a test you do not report.

Pass 6: Blind re-check

Re-read the flagged material once, deliberately without your diagnoses in mind. Fresh-eyes questions: where did attention drop? What confused? What did you skim? If the blind read does not reproduce the original symptoms, say so plainly and downgrade the diagnosis. The blind re-check is the loop's humility mechanism: it catches diagnoses the reviewer talked themselves into.

Pass 7: Verdict and revision task cards

Report:

• The surviving diagnoses (ranked by severity), each with claim, mechanism, evidence anchors, and what was falsified along the way.
• The killed diagnoses, with one line each on why they died. This is not filler. It shows the writer what was considered and ruled out, and it stops the next reviewer from re-running the same dead theory.
• One revision task card per surviving diagnosis (use assets/revision-task-card-template.md). Each card: the diagnosis, the exact scenes affected, the fix shape, a concrete first step, and a falsification note ("you'll know this worked when...").

Output Contract

Every review produced with this skill delivers:

1. Intake summary: what was read, the writer's stated intent, raw symptom list.
2. Diagnosis table: all candidates with claim, mechanism, fix shape, status (survived/killed), and anchors.
3. Falsification log: for each killed diagnosis, the counter-reading that killed it; for each survivor, the strongest objection it withstood.
4. Counterfactual results: test, passage, verdict, why.
5. Blind re-check note: what the fresh read confirmed or contradicted.
6. Revision task cards: one per surviving diagnosis, executable.
7. Sources: the rubric dimensions are grounded in the listed craft sources (see rubric file).

Operating Rules

1. Symptoms are not diagnoses. Never let "the middle felt slow" become the headline; it is the starting clue.
2. Minimum two competing diagnoses per review. One diagnosis is an opinion with a process costume on.
3. Every claim cites quoted text. No quote, no claim.
4. Never invent numbers, dates, quotes, or events. If the manuscript does not contain it, the review does not claim it. Unverifiable stays out or gets framed as the reviewer's read, explicitly labeled.
5. Review the story on its own terms. If you catch yourself prescribing a different story, kill that prescription.
6. Plain language throughout. No em dashes, no "delve", no "landscape" as metaphor, no "moreover", no "game-changer", no rhetorical-question setups, no motivational filler. If a sentence sounds like a press release, rewrite it.
7. When the loop kills everything, report that honestly. A review that says "the symptoms were misread, here is what I actually see" is more valuable than a confident wrong diagnosis.
8. Keep the review tight. Long reviews get skimmed; put detail in the cards.