Attention Review

Assess how a piece gives its audience a reason to start, a reason to continue, a reason to believe, a payoff worth the time, and something to take away. The method applies to any piece that has to hold an audience: an explainer, an educational or informational video, a podcast episode, a talk, a class, an article, a newsletter, a thread, a post, a presentation, or a story. These criteria guide editorial judgment. They cannot predict views, reach, or any individual's response, and a report must never claim they do.

This review sits on top of whatever structure the piece uses. It asks when the audience has a reason to keep going and when it receives what it was promised. It does not judge plot, character, or causality in fiction, shots or sound design in a production script, or the voice of the prose.

Scope and Expectations

Identify the material: a complete piece, a draft, an outline, an opening, or a published piece with its results. Record the medium, format, audience, purpose, platform, intended length, and stage when they are provided. If they are missing, form a brief hypothesis from the piece, state it under verdict and scope, and continue. Review a piece in the medium it has now: prose that may later be narrated is reviewed as prose, and the report notes that a narrated version would need its own review. Ask for the piece when there is nothing to review. A document's own heading counts as its working title. When the cover, thumbnail, or a title the piece will carry is not supplied, mark the checks that depend on it as not assessable and say so; do not stop to ask. When a stated target length and the actual content disagree, note it in the scope as a question; it becomes a finding only when it affects pacing.

Read the entire submitted unit before giving a diagnosis. For long pieces, proceed in sections and keep a compact record of every question the piece opens, where it is answered, and every promise made by the title, the opening, or the stated purpose. This record is a working tool; surface only the questions left unpaid, as findings. Verify final findings against the original passages. If only part was read, state the scope and present the diagnosis as partial. A fragment does not prove that a payoff is missing; mark that as not assessable.

Production scripts. Any script that carries cues, production notes, or markup is reviewed this way. Judge what the audience will hear and see at each point: the narration, the dialogue, the on-screen text, and the images the cues describe; when proof depends on what a cue shows, the cue is in scope. Metadata, production notes, and cue syntax are not the text. With markup such as SSML, strip the tags and keep the words: the spoken words inside the tags are the text. Do not flag markup as filler or count it toward length. When the script's own timing is missing or contradicts the narration, estimate durations from word count and marked pauses and label them as estimates. Do not evaluate framing, sound design, continuity, or production feasibility; if the user wants that review, the video-script skill covers it.

Paratext. These rules apply to any piece, prose included. Classification labels, ratings, genre tags, and framing notes such as a warning or "based on a true story" are packaging: judge them as part of the promise. A teaser for the next installment is transmission. Ignore credits and production lines such as suggested music or an estimated narration time; the target-length check above applies only to a length the piece is meant to meet. Author's notes appended to the piece, such as a list of planted seeds or a statement of intent, are not text: do not review them, but use them as the author's stated intent.

Stated intent. When the text works against an intent stated in the author's notes or in the request, report it as a finding conditioned on that intent. Type and rate it as if the intent holds, count it toward the statuses and the verdict on that basis, mark its correction as necessary while the intent holds, and offer both directions: change the text, or update the intent. Say in the verdict how it would read if the intent were updated instead.

Stories. When the piece is fiction, review only the attention layer: whether the opening gives a reason to read on, whether the genre promise is kept, whether open questions are paid off in steps, whether the rule or setup a resolution uses is on the page before the moment it pays off, whether the payoff lands and reframes what came before, and whether the ending leaves something to carry away. Whether the causal logic holds, whether a mystery's clues are fair, whether characters act from consistent motivation, whether events stay continuous from scene to scene, and plot holes are narrative questions. A contradiction the audience can see between two passages, such as figures or facts that disagree, is a proof finding even when it is also a continuity question; only the question of why it happened goes under "Outside this review". When a passage raises both, report its attention effect once as a finding, and add an "Outside this review" line only for the distinct narrative question, citing the finding's identifier instead of repeating its evidence; that is not double counting. A narrative question with no attention effect goes under "Outside this review" alone. Name each such question with its location and point to the narrative-review skill; do not diagnose it. An episode or chapter of a series may rely on setup from earlier installments the review cannot see; treat a resolution that could be prepared there as an interpretive risk and say so. Questions deliberately left open for later installments are transmission, not unpaid suspense, as long as the piece pays its own central question.

Hybrid pieces. A piece can tell a story and explain at once, such as a documentary-style fiction with a technical layer or an explainer built on an anecdote. Use the narrative reading for the story and the expository reading for the passages that explain, and say which applies where.

Performance data. Performance data is whatever shows how the audience actually responded: analytics, a retention curve, audience comments, reader or beta-reader reactions, the points where readers stopped, open or read-through rates. When the user supplies any, use it under Performance Data. When none is supplied, review the piece in full and add one closing line recommending that the user share the kind of data that fits the medium, if it exists. Never invent figures, and never delay the review to ask for data.

The Seven Stages

A piece holds attention when it delivers seven things, in the order its audience needs them:

1. Recognition: the audience quickly identifies a problem, desire, situation, or world it finds its own or can follow.
2. Gap: it discovers that something does not fit: its usual explanation is incomplete, or the situation contains a question without an answer.
3. Promise: it understands what it will get if it continues.
4. Proof: it receives reasons to believe: a demonstration, data, a case, a source, reasoning, or, in fiction, a coherent world whose rules hold.
5. Progress: it gets partial answers along the way, and each opens the next question.
6. Payoff: it understands something new, or sees what came before in a different light. This is the insight, the "aha".
7. Transmission: it knows what to do with what it received and has a reason to remember, apply, save, or tell someone.

The stages work on two axes. Recognition, proof, and payoff build understanding and credibility. Gap and promise trigger curiosity, progress sustains it, and payoff satisfies it. Curiosity behaves like a deprivation that arises from perceiving a specific gap in what one knows and fades when the gap is filled; it peaks when the audience believes it knows something but is unsure. The piece holds attention because it pays what it promises, not because it fosters a compulsive relationship with the content.

The order is flexible. A piece may show the proof before stating the gap, or open with the payoff and reconstruct how it was reached. A stage may not be missing: without a gap there is no reason to continue, without proof no reason to believe, and without a payoff the promise goes unpaid.

Checks by Stage

A "no" is a lead to verify, not a finding by itself.

Recognition. Would a member of the intended audience recognize the subject in the first stretch? Is the problem expressed in the audience's words? Is the scene specific, or could it belong to any audience? Does the opening match what the title, cover, thumbnail, or announcement promise? In fiction, recognition does not require an everyday world; it requires a legible one, usually through a character who wants something we understand.

Gap. Is the gap concrete? "You won't believe this" creates expectation but does not say what is missing or why it is worth knowing. Working gaps take the form of a contradiction, a hidden cause, a comparison, a prediction the audience is invited to make, a specific question, or, in fiction, an anomaly or an inciting incident. Packaging that gives away the answer closes the gap before the piece opens it; a genre word in a title ("ghost", "murder") is a promise, not a spoiler.

Promise. Is it clear without extra context, relevant to the audience, proportional to what the piece can deliver, specific rather than "secret" or "ultimate" with nothing behind them, and near? Near means the first partial answer arrives before the audience's first likely exit point: the first seconds in a feed, the first section of a long video, article, or talk, the first scene of a story. Relabeling at the end, as the promised "secret", an answer already given midway is a promise mismatch, not a payoff. In fiction, genre and tone make the promise; breaking it without offering something better reads as a swindle.

Proof. Is authority demonstrated inside the piece rather than claimed through titles or years of experience? A credential alone has at best a small effect on persuasion. A credential is fine as context once the proof is shown; it is a finding when it replaces the proof. An invitation to try it at home complements proof shown in the piece and never replaces it. Useful sources of proof are demonstration, data with context, a comparison of equivalent cases, direct experience with its conditions, a named external source, and reasoning without hidden leaps. Trust signals: observation kept apart from interpretation, the limits of an example acknowledged, correlation not presented as cause, sources shown when a claim can be checked, the conditions under which advice works stated. In fiction, proof is verisimilitude: the world's rules stay consistent, and the audience receives a rule or a means on the page before it resolves anything. Do not rule on the domain truth of a claim beyond what can be checked: figures that contradict each other or a range the piece itself states are proof findings, and so is a claim that contradicts well-established knowledge once the piece's own figures are put side by side; a claim you merely have reason to doubt is a low-confidence finding that recommends verification. When the result of a demonstration cannot reach the audience through the medium, such as a taste or a smell in a video, the piece must translate it into something perceivable, such as a reaction, a measurement, or a visible comparison; otherwise the proof is only asserted. When such a defect originates in source material outside the piece, such as a series bible or a reference document, report it where the audience meets it and name the likely source to fix.

Progress. Does the piece work in short loops of question, useful partial answer, and next question, closing the main idea before the end? A single promise at the start loses force. There is no universal interval; progress should arrive when the path needs it, not every fixed number of seconds or paragraphs. Rapid cutting, short sentences, or visual stimulation renew attention but do not substitute for an idea that advances. In fiction, each loop is a scene that changes something and narrows the open question without closing it.

Payoff. Does it resolve the question asked, compress several observations into a clear rule, surprise without being arbitrary, feel coherent once seen, and allow a decision or an action? Insights experienced as an "aha" are recalled better later. Apply the insight test: "I used to think ___. Now I understand ___, because ___." If the blanks cannot be filled precisely, the piece offers information but no clear payoff; in fiction, an unfillable "because" means the reframe does not land, and whether it was fairly prepared goes to "Outside this review". Weak payoffs restate the promise, reveal a fact that changes no decision or reading, present an opinion as a discovery, reserve all usefulness for a second part, or stay too general to apply.

Transmission. Is there a reason to save (steps to use later, a checklist, a summary of something hard, a template, a reference) or to share (it helps someone with a recognizable problem, expresses an identity or stance, settles a common argument, warns of a costly mistake, produces surprise, humor, or admiration, or opens a relevant conversation)? Does the close match the value delivered, and does it ask for one thing? "Follow me for more tips" is disconnected; "save this and use the seven stages as a checklist before you record" and "send it to whoever still tries to fix retention with faster cuts" are tied to the value. In fiction there is no call to action; transmission is resonance, and it works when the theme emerges from what the characters did rather than from a closing speech. Calls made inside the story are story content. A direct address that asks the reader to act in the real world is judged as a call to action.

Two Readings of the Same Stages

In a story, the order of information produces the effect: showing the initiating event and delaying the outcome creates suspense, showing the outcome before its causes creates curiosity, and withholding a key fact and revealing it suddenly creates surprise. Gap and promise decide which of the three is active; the payoff is where it resolves.

Balance

A strong piece combines the familiar and the new. Repeated exposure raises liking only up to a point, and what is easy to process tends to please, which is why the familiar opens the door; what is entirely known gives no reason to stay.

• Everything familiar: predictable.
• Everything new: too much effort to follow.
• Promising more than it delivers: loses credibility.
• Explaining without opening a gap: useful but not compelling.

The working formula is something the audience knows, plus something that does not fit, plus an explanation it can check, plus an application or a consequence.

Humor

In humor the payoff changes function: the reward is an unexpected interpretation that produces laughter, surprise, or recognition. The adapted structure is recognition, expectation, turn, punchline, and social identification.

Authority weighs less; comic voice, timing, performance, and character consistency weigh more. When the humor depends on a profession, hobby, or community, knowing its details does work as authority, because it makes the parody precise for insiders.

Checks: the situation is recognized without a long introduction; a clear expectation exists before the turn; everything needed appears before the punchline; nothing remains that does not set up, escalate, or pay off the joke; the turn surprises and is understood at once, growing from a word, a character, or a rule already established; timing leaves room for the punchline to land; each tag (a further punchline on the same premise) escalates, varies, or closes the joke rather than repeating it; the ending is stronger than the middle jokes; each attempt in a sketch changes the situation, since the same reaction louder is not escalation; the audience can think of a specific person; the joke does not depend on a false claim presented as fact.

Common failures: explaining the joke; hiding the premise so long that no expectation forms; mistaking randomness for surprise; running past the best punchline; copying a format with no observation or point of view of its own; attacking a vulnerable person instead of observing the contradiction or absurd logic that sustains the premise.

Feed Video

Apply this section only when the piece is a short vertical video for a feed, such as TikTok, Instagram Reels, or YouTube Shorts. In a feed the platform places the video in front of someone who did not ask for it, and the decision is to keep watching or swipe away. That moves the weight of every stage toward the first seconds. Platforms state that their recommendations rely on signals of whether people choose to watch, how long they keep watching, whether they are satisfied, and what they do next, such as saving, sharing, or following.

Add these checks to the stage checks:

• First frame and opening text. In the feed the first frame is the cover. Do the first frame, the opening on-screen text, and the first line say the same thing, and do they carry the subject and the promise?
• With and without sound. No reliable current figure exists for how many people watch short vertical video muted; the widely quoted numbers are old and come from other feeds. Design for both: the hook, the gap, and the proof must be readable on screen, and sound should still add rather than be sacrificed.
• Safe zones. Each app's interface covers part of the frame, typically the top, the bottom, and the right edge. No platform publishes verifiable official dimensions, so the check is practical: text, proof, and faces must stay clear of the interface in the app's own preview. From a script, this is a finding only when the script places something important in a covered zone; otherwise mark it not assessable and recommend the preview check.
• Early proof. The proof appears in the first stretch, visible at the moment the promise is made.
• Payoff with margin. The insight lands clearly before the end; whoever swipes before the last second never receives a payoff placed there.
• Loop. If the last image or line connects to the first, the replay comes naturally. Check that the connection does not leave the promise unclosed: a replay caused by confusion is not the same as one caused by enjoyment.
• Duration. The idea decides the length, not the platform maximum. Limits change; as verified on 2026-09-22, YouTube Shorts accepts up to three minutes and Instagram recommends reels to non-followers only up to three minutes. When the piece approaches a limit, tell the user to confirm the current one before relying on it.
• One action. The call to action asks for one thing tied to the value delivered. Following makes sense when the profile promises a continuity the video has just demonstrated.
• Reposts. Instagram states that it recommends reposted content carrying another platform's watermark less often. Flag a watermark only when the exported video is in view; a script cannot show one, so omit this check for scripts.

For humor in a feed: keep the punchline near the end and the end near the punchline, because the stretch after the best punchline is where most viewers leave; a closing callback reads with new meaning on the replay; social identification has a button, since the viewer can tag or send it in one gesture.

Performance Data

Use this section only when data is supplied. Map each symptom to the stage that most plausibly explains it, and present the mapping as a hypothesis to test, not a verdict. The table below is written for published video and posts. For text, the same logic applies to where readers stop, what they quote, and what they ask: map each point to the stage the piece was in at that moment.

Cautions when reading data:

• Views are an end result shaped by the content, audience size, topic, competition, timing, and distribution. Diagnose each stage with a related metric instead.
• Replays can push retention figures up; separate first-view retention from replays before comparing.
• Platforms change how they count views. If the counting method changed within the compared period, the earlier series is not a baseline.
• Prefer ratios, such as shares per reach, to absolute counts, and the median of a series to a single video.
• A comparison is informative when one main element changed (opening, placement of the proof, form of the insight, call to action) across pieces of similar topic, format, and length. One piece is not a series.
• Do not apply universal thresholds; a healthy percentage depends on platform, length, niche, traffic source, and audience.

Ethical Limits

Report as findings any use of attention techniques to hide a weak payoff or a false claim:

• A promise the piece does not keep.
• Suspense that is never paid, or paid only at the end to hold the audience. An open question is legitimate when the answer arrives reasonably soon and delivers what was promised; questions left open for later installments of a series follow the rule under Stories.
• Data without context that exaggerates a result.
• Credentials irrelevant to the subject.
• Invented urgency or fear, or anger provoked so the piece spreads. High-arousal emotions such as anger and anxiety do increase sharing; that is the reason to refuse them as a device, not to recommend them.
• Medical, legal, financial, or security advice without adequate sources and limits.
• Editing that changes the meaning of proof or testimony.
• In nonfiction, fictional devices the audience cannot distinguish from fact.
• In fiction framed as true, such as "based on real events", a found archive, or a mock documentary: the frame is a genre convention when it is clear or openly ambiguous. It becomes a finding when the piece hands the audience real-world advice or factual claims it could act on without marking them as part of the fiction.

Prioritize an ethical finding by the harm it can cause: high when the audience could act on it to its own detriment or it misleads about the piece's main claim, medium when it misleads about a secondary point, low for a minor overstatement. Name "ethics" as its stage, followed by the stage it damages.

A related trap: the feeling of insight also accompanies wrong answers, though usually more weakly. A payoff that feels right is not thereby true; when a surprising claim lacks sufficient evidence, the finding is to weaken the claim or add proof.

Verifying Findings

For each potential problem:

1. Locate the passage that produces it: a section and timestamp, verified line numbers, or a short identifiable quotation. Do not invent timestamps, line numbers, or quotations.
2. Name the stage affected and explain how the problem may shape the audience's experience at that point.
3. Look for a supported alternative explanation: a later payoff, a deliberate delay that is paid soon, genre intent, a platform convention, or material outside the submitted fragment.
4. Decide its type: a supported problem, an interpretive risk, or an editorial preference. The type rests on the evidence for the defect, not on its effect, since an effect on the audience is always a reasoned interpretation: it is a supported problem when the text itself shows the defect, such as a repetition, a contradiction, a delay that can be measured, or a promise left unpaid; it is an interpretive risk when the passage is a defect only under one reading of it, including when a visible pattern such as a repetition has a supported explanation from step 3, like a genre convention or a stated intent that requires the pattern itself. An intent that justifies the scene containing the pattern, but not the pattern, leaves it a supported problem. A supported problem calls for a necessary correction; an interpretive risk or an editorial preference calls for an optional recommendation. If the evidence is insufficient, discard it or leave one concrete question.
5. Propose the smallest intervention that addresses the cause, such as moving the proof before the introduction, turning a vague promise into a specific one, closing a loop that was left open, or cutting the stretch after the payoff. State its cost when it changes the structure, the tone, or the ending.

Describe audience effects as reasoned interpretations, not facts about viewers. Group symptoms that share a cause while keeping their separate locations, and do not count one problem twice. A grouped finding lists every stage it touches, the stage where the cause sits first.

Prioritize by impact:

• High: compromises the main promise, the main payoff, or the reason to continue in the opening. An opening that slows down after the gap is already open is not a failure of the opening; rate it by how much it weakens the stage.
• Medium: appreciably weakens a stage, a loop, the proof, or the close.
• Low: a localized problem whose fix improves clarity, economy, or timing.

The central question is the one the opening raises and the piece commits to answer; in an installment of a series, the one this installment answers. When the title promises something else, the title's promise and the central question are both main promises. The main payoff is the one that answers them; in fiction, the emotional close that follows it belongs to transmission. A finding that damages only a secondary question weakens its stage rather than keeping it from working.

Rate confidence separately:

• High: the passage shows the problem directly.
• Medium: the evidence supports it, but a plausible alternative remains.
• Low: a suspicion worth checking.

Keep priority separate from confidence: a serious suspicion with little evidence remains a suspicion. Do not invent findings to fill a quota. A piece may hold attention without requiring corrections.

Report

Write the report in the user's language unless requested otherwise. Match its length to the piece and the assignment. For a general review, include:

1. Verdict and scope. State whether the piece holds, partly holds, or loses attention, and where. Decide it from the statuses in the table, applying the first that fits: it loses attention when recognition or gap has the status "supported problem" so that the opening gives no reason to continue, or when promise or payoff has that status; it partly holds when any other stage has that status, when payoff is "partial", or when any medium- or high-priority finding calls for a necessary correction; it holds otherwise, with any partial stages and remaining corrections named. A delayed opening that still gives a reason to continue does not lose attention. Name the material reviewed, the inferred medium, format, platform, audience, and purpose, and whether the feed checks applied.
2. Strengths to preserve. Effective mechanisms with brief evidence, especially those a correction might damage.

3. Status by stage. A table with each of the seven stages, its status, and its basis. The statuses, translated into the report's language:

   • Meets expectations: the stage does its job; at most optional notes or a low-priority correction confined to one passage or to one contradiction between two passages.
   • Partial: the stage works, but a finding weakens it, or its only findings are interpretive risks. A low-priority finding of any type confined to one passage or to one contradiction between two passages, or an editorial preference, leaves the stage at "meets expectations" with the note.
   • Supported problem: a supported problem keeps the stage from doing its job.
   • Not applicable or not assessable from the available material.

   A finding that touches several stages sets "supported problem" on each stage it keeps from working and "partial" on each it only weakens. When the feed section applies, add rows for its checks; when there is an ethical finding, add an ethics row. Do not use numerical scores unless the Score subsection applies.

4. Prioritized findings. Number findings A1, A2, and so on in one sequence, in priority order. For each high or medium finding: identifier, priority, type, confidence, location, stage, evidence, mechanism, impact, and a concrete proposal marked as a necessary correction or an optional recommendation. A low finding may be compact: identifier, priority, type, stage, location, evidence, and a proposal marked as necessary or optional, adding confidence when it is not high.
5. Revision plan. When one intervention depends on another, the dependency sets the order: a conflict with the stated intent comes first, because its direction decides the rest; the alignment between gap, promise, and payoff comes before proof and progress, which come before the opening and the close, which come before feed checks. Where no such dependency exists, order by priority. A finding that is both an opening and a feed issue goes with the opening; an ethical finding goes with the stage it damages. For each major change, state how to verify that it works and what to preserve. When nothing needs a necessary correction, say so and list the optional recommendations briefly.
6. Outside this review. One line per symptom that belongs to another kind of review (plot or character, production, prose voice, copyediting), with its location and the skill that covers it, or a note that no available skill does. Prose-voice symptoms, such as stiff or redundant phrasing, repeated or near-identical lines, or a register that slips, belong to the humanize pass that runs over the whole text when corrections are applied; list them with their locations, and say which repetitions look deliberate so the style instruction protects them. Spelling, grammar, punctuation, and typography errors belong to the proofread skill, which runs once the text is final: give them one line naming the kind of error and a few locations, not a line each. Omit the section when there is nothing to list.

When no performance data was supplied, end the report, after the last section, with one line recommending that the user share it if it exists, saying which symptom it would help confirm. Do not provide a long summary of the piece unless it is needed to explain a finding. For focused assignments, such as "only the opening", develop only the requested area and the dependencies that affect it.

Score

Compute a score only when the user asks for one, or when the user supplies two versions of the same piece to compare. Score each criterion 0 (absent), 1 (present but weak or ambiguous), or 2 (clear, specific, and well executed), and show every criterion with its score and a one-line basis, never the total alone:

1. The audience recognizes the subject or world immediately.
2. The opening matches what the title, cover, or announcement promise.
3. There is a concrete gap.
4. The promise can be kept within the piece.
5. There is verifiable proof, or sufficient rules and setup in fiction.
6. The path delivers progress before the close.
7. The payoff changes or clarifies understanding.
8. The idea can be applied, or the consequence resonates.
9. There is a natural reason to remember, save, or share.
10. The close asks for one thing, or for nothing.
11. There are no misleading or exaggerated claims.

The maximum is 22. When the feed section applies, add six criteria, for a combined maximum of 34: first frame and opening text carry the subject and promise; proof appears in the first stretch; the video is understood without sound; text, proof, and faces stay clear of the interface; the payoff lands with margin before the end; the duration is what the idea needs. For humor, replace criteria 3 to 5 and 7 to 8 with the humor checks, each scored 0 to 2. Mark a criterion that cannot be assessed from the available material as not assessable, leave it out of the total, and state the resulting maximum.

State that the score compares versions of the same piece and does not predict results. When comparing two versions, report the score of each, the criteria that changed, and whether any change solved one stage by breaking another.

A review authorizes diagnosis and recommendations. Modify the piece only when the user requests corrections, following Applying Corrections. You may include a short example labeled as a proposal when it clarifies an intervention; never present it as a quotation from the original. For a second review, verify the earlier findings against the new version and state which are resolved, remain, or changed.

Applying Corrections

Rewrite only when the user asks for it, and only the findings the user selected. Where a finding offered options, have the user choose before touching the text.

Edit only the passages cited in those findings, including the place where a proposal says the fix goes. In this first stage everything else stays byte for byte, including markup, cues, production notes in a script, and line endings. Do not add sections or scenes unless the chosen option requires one. An inserted line imitates its neighbors: sentence length, vocabulary, tense, register, and, in a script, the spoken rhythm of the narrator. It delivers what the finding needs without explaining it, and it carries no filler connectives or "not X but Y" frames. When a correction moves or adds proof, its factual content is fixed: the wording may change, the facts may not, and no new data may be invented to fill a gap; if the proof does not exist, say what the user must supply.

If the humanize skill is available, run it in two places. First on the inserted lines, supplying the rest of the text as protected material so it can read the author's voice and imitate it. Then, once the findings are applied, run it over the whole text as a second stage of the correction: prose written by a person can still be stiff, redundant, or unclear. Give humanize a style instruction built from the review: the genre and register, the strengths to preserve the report named, and the deliberate devices that are style rather than defects, such as anaphora and parallel negations, contrasts of the form "it was not X; it was Y", deliberate fragments and one-word paragraphs, escalation by repetition of a word, and, in Spanish, the raya that opens dialogue or sets off a narrator's inciso and the ellipsis character …. Tell humanize that the style instruction overrides its own defaults on these points. Supply the spoken words of dialogue, including dialogue the correction inserted, quoted material, and, in a script, markup, cues, and production notes as protected text. Leave headings, epigraphs, the author's notes, and documents quoted inside the story, such as a letter or a diary entry, outside the pass. In a script, the words a listener hears inside markup may be edited like any narration. The pass may change rhythm, word choice, and clarity; it never changes facts, names, figures, events, their order, or what a character says, and it keeps the file's line endings. After it, check that every device named in the style instruction survived, and revert any edit that removed one. When humanize is not available, say so in the change log. Both passes run by default. Skip them only when the user explicitly asks, for instance by saying not to humanize or that humanize will run later in their own workflow; never infer it from the size of the correction or from context. An explicit request skips both passes, the inserted lines included: the inserted lines then rely on imitating their neighbors, as the drafting rule above requires. When humanize is skipped, say in the change log that it was skipped at the user's request, and write "none" with that reason in its humanize part. When the user asks to change little, for instance by saying not to change much, the passes still run but are limited to errors and clarity problems, such as contradictions, broken constructions, and obvious redundancies; they leave rhythm and style alone, and the change log says the limit was applied.

Deliver the complete corrected piece and a change log in two parts: one entry per finding with identifier, location, text before, text after, and what to verify; then one entry per edit the humanize pass made: location, text before, and text after, so each can be reverted on its own. Give every location in the original's numbering, and locate an insertion as "after Lnn". Then review the new version as a second review: state which of the applied findings are resolved, remain, or changed, mark the findings the user did not select as not selected, and confirm that everything listed under strengths to preserve is intact.

Example Use

"Use /attention-review on this explainer script for YouTube Shorts. Tell me where viewers are likely to leave, whether the proof arrives in time, and what to change in the opening. Here is the retention curve from the last version."