---
name: cutkit-script-check
description: Write short-form video scripts (TikTok, Reels, Shorts) that pass CutKit's script check. Use when someone asks you to write, draft, tighten or fix a hook, opening line, short video script or talking-head script, or asks if a script is any good, and CutKit is connected. Covers the 20 checks, the hook shapes with an example each, the length limit per platform, and how to call the check tool before you save the script.
---

# CutKit script check

CutKit checks a script's craft before the person films it. It doesn't predict views. A model reads the script and answers 16 questions. The other 4 need no model: a count of the first line's words, a scan for hedge words, a scan for long stretches with no turn, and a scan for AI-sounding text. That's 20 checks, and the answer gives a `score` out of 100. The score weighs the checks. It's a craft count and not a forecast.

**Craft and topic fail on their own.** Three of the questions judge the writing itself: `coherence`, `voice` and `specificity`, each scored 0 to 3 and passing only at 3. So keep rewriting the words until they pass. Don't tweak the script to tick boxes: a named idea, stakes, a re-hook and a CTA on incoherent prose still fail `coherence`. **A boring topic fails too.** `topic_pull` judges the subject and angle, not the wording, scored 0 to 3 and passing at 2 or 3. Internal admin, tool tidiness and process hygiene with no money, status, fear or story attached score low however well they're written. Rewording won't lift that. Pick a different angle with the person: a cost, a result, a mistake or a story. Write to these checks first, then run the check, then fix what fails.

This sits beside the main [cutkit skill](https://getcutkit.com/skills/cutkit/SKILL.md), which covers connecting and saving scripts. Read that one for auth and the `create_script` tool.

## The loop

1. Ask only what you can't guess: the subject, who it's for, the platform.
2. Draft to the rules below, or start from a template for the hook: pick one with `list_templates`, then `start_from_template` saves a draft with a `[bracket]` for every beat. Read it at speaking pace. For a story-led script, `start_story` saves the four story beats instead.
3. **Lint until clean, then spend a real check.** Call `lint_script` (REST `POST /api/v1/lint`, CLI `cutkit-scripts lint --file F`). It runs only the 5 offline rules (first line, greeting opener, hedges, flat stretch, AI-sounding text), costs no quota and needs no model. Fix what it names, lint again, and only then call the check.
4. Call the check. MCP tool `check_script` with `body` (the script words), or `id` or `title` (a saved script). REST: `POST https://getcutkit.com/api/v1/check` with `{"body": "..."}` or `{"id": "..."}` and `Authorization: Bearer ck_...`. Same key as the scripts API. CLI: `cutkit-scripts check <id or title>`, `--file <path>` or `-` for stdin, and `--json` for the raw result.
5. Read the score first. The MCP text and the CLI then list each failing item with its fix and the templates that fit. For every item with `pass: false` that isn't `unsure`, apply its fix (table below). When it lists `templateIds`, read one with `get_template` and borrow its shape.
6. Save with `update_script`, then check again. Keep going until nothing that isn't `unsure` still fails, or until the failures left are ones you can't fix without inventing facts. Tell the person the score and what still fails and why.

The answer's fix text for `no_slop` quotes up to 4 offenders. Fill every template blank before you check, because a leftover `[like this]` fails it.

Checking is read only and saves nothing. It needs 15 to 2,000 words and allows 100 checks a day per library, so draft well before you spend one. Every answer says how many are left today (`quota`) and when the next slot opens. Checking the same text again within 24 hours returns the stored answer and is free (`cached: true`), so re-checking identical text never costs anything. It can answer `check_unavailable` (503): wait a moment and retry. It's a model, so a close call can flip between runs. Items carry a `confidence`, and under 0.6 the item is `unsure`: a yes-or-no item that's unsure counts 0 in the score, the 0-3 checks (`opening_speed`, `coherence`, `voice`, `specificity`, `topic_pull`, `audience_fit`) still count at their level, and the answer lists unsure items apart from the fixes, so don't rewrite a good script over it.

## From a topic to a clean check

When the person gives you only a topic, run this end to end and ask nothing else. Write the words yourself and use CutKit only to check them. `rewrite` is for people without an agent.

1. **Test the topic first.** Before writing, ask whether a stranger would stop for it: a cost, a result, a mistake or a story attached? If it's housekeeping or a how-to nobody looks up until they hurt, propose a sharper angle to the person before you write a word. A boring topic fails `topic_pull` whatever you write.

   Then **pick the family from the topic.** Decide what the viewer is stuck on, then take the shape that answers it.

   | The topic is... | Family |
   | --- | --- |
   | a result they want fast | `promise` |
   | something not working, cause unknown | `diagnosis` |
   | advice everyone repeats that you can show is wrong | `belief-break` |
   | a moment you were in | `in-medias-res` |
   | one kind of person with one fear | `tribe` |
   | something they can open and look at | `artefact` |
   | a test, a count or a measured result | `number` |
   | a question they already search for | `question` |
   | something that just changed | `news` |
   | two true things that don't fit together | `contradiction` |
   | a mistake that costs them | `warning` |

2. Call `list_templates` with that `family` and take the one template whose blanks you can fill with true facts. Never take a shape you'd have to invent proof for.
3. `start_from_template` with its id. Write real words over every `[bracket]`, keep the opening shape and the re-hook, then `update_script`.
4. `lint_script` until clean, then `check_script` by `id`. Fix every failing item, heaviest weight first, save, check again.
5. **Stop** when no item that isn't `unsure` still fails. When a round doesn't raise the score, switch family rather than polish the same one. Stop after 5 rounds, or when what's left would need an invented fact. Then tell the person the score, the hook id you used and what still fails.

## The 20 checks

The answer lists these as `items`, each with an `id`, `pass`, `fix` and `templateIds`. The score weights them, and specificity counts most: `specificity` 6, then 2 each for `opening_speed`, `hook_shape`, `rehook`, `one_idea` and `strong_close`, and 1 each for `contrast`, `topic_pull` and `audience_fit`. Each counts at its share, not just pass or fail: a 2 out of 3 on specificity earns two thirds of its 6. The other twelve count 0. Across 4,527 TikTok scripts, they added nothing once specificity was in, so they still show a fix but don't move the score. `coherence` and `voice` count 0 in the score but still fail with a fix. Fix the heavy ones first.

| id | Passes when | If it fails |
| --- | --- | --- |
| `hook_length` | The first sentence is 15 words or fewer. | Cut the first line to one claim of about 10 words. |
| `opening_speed` | The first 2 sentences name the subject and what the viewer gets (a score of 2 or more out of 3). A greeting, a line that hides the point ("wait till you see this") or one that opens on a time and place fails it outright. | Say the subject and the result in sentence one. Cut the greeting. Put the time and place later. |
| `hook_shape` | The first sentence matches one of the 11 hook shapes below. | Rewrite the opener into one of the shapes. |
| `contrast` | The first 3 sentences set what the viewer believes or does now against something different, said out loud or implied ("8 times faster"). | Say what most people think or do, then what you'll show instead. |
| `fresh_angle` | Most people the script is for would hear something they didn't know: an uncommon idea, or a common one from an unusual angle. | Lead with the part they haven't heard, or find an angle nobody uses. |
| `second_beat` | Sentence 2 or 3 answers the doubt a viewer has about the opening, or shows proof. When the opening flips a belief, the belief is stated out loud first. | Name the doubt, then answer it, or rule out the easy answer. |
| `rehook` | About halfway, there's a new question, a higher stake or a revealed mechanism. "This is where it gets crazy" doesn't count. | Add a fact, a question or a flip in the middle. |
| `payoff` | The script delivers what its first 2 sentences promise. | Show the thing you promised, or change the promise. |
| `ask` | One ask, placed at the very end, right after the best point, or as a one-word comment to get something. No ask, or several at once (comment, follow and share), fails. | One thing to comment or try, after your best point. |
| `named_idea` | The script gives its main idea or method a name of its own. | Your best idea doesn't have a name yet. Call it something people can repeat. |
| `stakes` | The script names who it's about, what they could lose, and why now. | Give it stakes: who, what they could lose, and why now. |
| `one_idea` | The script follows one main idea from start to end. | Keep the strongest idea. The other one is its own video. |
| `strong_close` | Leaving out the ask, the last line is short, specific and about as strong as the first, a line a viewer would quote back. A summary or sign-off fails. | Write the last line first. A short loops, so it plays right before your hook. |
| `no_hedges` | The first 2 sentences hold no hedge (if, maybe, might, could, probably, perhaps, possibly, hopefully, kind of, sort of, I think, I guess), and the whole script stays under one hedge per 100 words. | Say it as fact. "If you try" becomes "when you try". |
| `loop_gap` | A script over 120 words never runs more than 70 words without a contrast word (but, actually, turns out, except, however, yet, though, instead, until). | Nothing turns for too long. Add a "but" or a question there. |
| `coherence` | Every sentence follows from the last. No non sequitur, no dangling "Yet" or "But" with nothing to contrast, no line said twice or restated in other words. Level 3 of 3 to pass. | Read it aloud. Cut the repeated line, and make each sentence lead to the next. |
| `voice` | Reads like one person talking to one person: contractions, concrete nouns and numbers, short and long sentences mixed. A flat, abstract or templated script fails. Level 3 of 3. | Say it to one person. Add contractions, swap abstract nouns for real ones, break up equal length lines. |
| `specificity` | Every claim carries a real object, number, name or moment. "Processes", "systems" and "workflows" with nothing behind them fail. Level 3 of 3. | Swap each generic word for the thing itself: the tool, the number, the day it happened. |
| `topic_pull` | The subject, whatever the wording, is a recognizable pain or desire with a visible cost, or a concrete result, for a named audience (level 2), or something a viewer would send a friend: a surprising number, a costly mistake, a claim with proof, a story with an outcome (level 3). Housekeeping and look-it-up reference fail. | Change the subject or find the angle with a cost, a result or a story. Rewording won't fix this. |
| `no_slop` | Nothing reads as written by AI: no `[bracket]` left over from a template, no long dashes, no emoji, and no pile-up of stock AI words and phrases: one in every 100 words, and at least 2. A lone "I think" or "genuinely" passes, because people say them. | The fix names what it found. Rewrite those parts the way the person would say them. |

### Greetings that fail the opener

An opening that starts with any of these fails `opening_speed` with no model involved: hey, hi, hello, welcome, "so basically", "in this video", "today we're going to", "today I'm going to", "let me tell you". So does a line that hides the point ("wait till you see this", "you're never going to believe this") or opens on a time and place ("last month I was at...", "it was a cold night in..."). Start on the point.

## Hook shapes

`hook_shape` wants the first sentence to be one of these. One example per shape. The check's `openingFamily` tells you which one it saw.

| Shape | What it does | Example |
| --- | --- | --- |
| `promise` | Promises a specific result or change. | This is the only pan you need to cook dinner for two. |
| `diagnosis` | Names a problem and its cause. | Your basil hasn't grown because the water hits the leaves and misses the soil. |
| `belief-break` | States a common belief, then contradicts it. | Everyone will tell you to post every day. Your best customers may only need one useful note a week. |
| `in-medias-res` | Starts inside a concrete moment or action. | You just sent the invoice, and now the total looks wrong. |
| `tribe` | Calls out a specific group or identity. | If you want to squat but you're afraid of your back, brace first. |
| `artefact` | Opens with a visible object, result or proof. | Open your calendar. Look at how many meetings have no decision attached. |
| `number` | Uses a specific number as the reason to listen. | I spent 3 mornings testing porridge ratios so you can stop scraping the pot. |
| `question` | Asks a real question the viewer wants answered. | Does cold water actually keep your pasta from sticking? |
| `news` | Leads with a new event or change and why it matters. | Your bank just released card locks, so you can now freeze a lost card. |
| `contradiction` | Puts 2 facts in direct tension, naming the common belief first. | Your wooden spoon is more useful than it looks because its handle measures one portion. |
| `warning` | Names a pain, habit or mistake the viewer should avoid. | The worst thing you can do for basil is water the leaves. |

Keep the first line to 15 words or fewer. Several of these examples run 2 sentences; only the first sentence counts for `hook_length`.

## Re-hook moves

For `second_beat` and `rehook`. Every template in the library pairs its opening line with a re-hook, so a failed `rehook` lists `templateIds` you can look up and lift the middle line from.

| Move | Example |
| --- | --- |
| Name the doubt | You're probably asking what happens when the pan is already hot. |
| Not A or B | It isn't the flour or the oven. It's the warm butter. |
| Result first | This is what you get: crisp edges and a soft middle. Now dry the potatoes. |
| Better way | But here's a better way to peel garlic: press the clove first. |
| Condition | But this won't work unless you let the pan heat first. |
| Flip | You'd expect the larger pan to cook faster. But the crowded burner leaves its edge cold. |
| Count | There are 3 parts. The first is the water already on the mushrooms. |

For the doubt right after the hook, naming the doubt and ruling out the easy answer are the fastest. The MCP tool `list_templates` (REST `GET /api/v1/templates`, CLI `cutkit-scripts templates`) returns every template with its id, its family, its opening line and its re-hook. Narrow it with `family` set to one of the hook shapes above. `get_template` (REST `GET /api/v1/templates/{id}`) gives one in full, with 2 examples and the way it goes wrong.

## Review promises and open loops

This is an agent's editorial review, separate from the 20 API checks and their score. `payoff` covers the opening promise; `rehook` looks for a middle turn. Neither result certifies that every later question gets answered.

For a script whose beats depend on each other, try: hook and concrete promise, useful content beat, specific unresolved question, next beat answering it, final payoff and one relevant ask. The hook can contain the promise; the script needn't have 3 parts or 2 loops. Source: [Loïc Bourget's reel](https://www.instagram.com/reels/DdmLvdTIdyR/), inspected 5 Oct 2026; a creator's structure, not controlled evidence of conversion.

Keep a promise-to-payoff map in the chat or brief, outside the teleprompter body. Review the finished words:

1. Each promised result and open question has a later line or planned visible proof that resolves it before the video ends.
2. Each beat delivers a useful fact, example or action. A transition names the next specific unknown; "the next part is even crazier" doesn't pass. Removing suspense language leaves a coherent explanation.
3. The single ask helps the viewer apply the delivered result. Any offered resource exists; a comment or DM cannot be required to receive an answer the video promised.

Fix failures in at most 2 editorial passes. Remove or narrow promises that require invented evidence and report remaining gaps separately from the API score. Keep mapping and filming notes outside the spoken body. A high score doesn't waive this review or establish retention or conversion.

## Length

The check reads the script aloud at 140 words a minute, rounds up, and compares the seconds to each platform. The answer's `platforms` list says which fit.

| Platform | Limit | Words at most |
| --- | --- | --- |
| `tiktok` | 60 seconds | 140 |
| `reels` | 90 seconds | 210 |
| `shorts` | 60 seconds | 140 |

These are starting values, not measured data. Over the limit, the note says to cut a beat unless every one is new. When the person names a platform, write to its number. When they don't, aim for 140 words or fewer so it fits all three.

## Before you save

- First sentence is 15 words or fewer, with no greeting, and is one of the 11 shapes.
- The subject and the result are both in the first 2 sentences, set against what the viewer believes now.
- The idea is one they haven't heard, or a familiar one from a new angle, and it's the only idea.
- Each sentence follows from the one before, nothing repeats, and every claim has a real object, number or moment behind it. Craft fails on its own, so this comes first.
- The last line is short enough to quote.
- Sentence 2 or 3 names the doubt and answers it.
- Halfway through, something new: a fact, a question, a flip.
- The opening promise and every later open question get resolved before the end; the editorial review above passes.
- Nothing says the video is worth watching.
- No hedges in the first 2 sentences, and the script says it as fact.
- A long script turns at least once every 70 words.
- The script gives two useful points before it asks, names its best idea, and gives the stakes: who, what they could lose, and why now.
- No `[blanks]`, no long dashes, no emoji, no stock AI phrasing, and more "you" than "I".
- One ask, after the best point or at the very end, tied to applying the result; any offered resource exists.
- Under the word count for the platform.
- Everything is a spoken word. No `[pause]`, no stage directions, no markdown. The main skill has the teleprompter rules.
- You ran `check_script` on the final text, and not on an earlier draft. Tell the person its score.

Don't claim the script will do well. The check says whether the craft is there and nothing about views. If an item still fails, say which and why.

## More

- Main skill: https://getcutkit.com/skills/cutkit/SKILL.md
- Docs: https://getcutkit.com/docs/api
- OpenAPI 3.1 spec: https://getcutkit.com/openapi.json
