The five parts of a controlled prompt
A one-line request such as "rewrite this for LinkedIn" leaves the model to decide what counts as a fact, what counts as style, and what a good post looks like. It fills those gaps with the patterns it has seen most: a dramatic hook, a personal anecdote, emojis, hashtags. A controlled prompt answers those questions before the model has to guess.
| Part | What to write | What it prevents |
|---|---|---|
| Source | The original post, labelled as the only source of facts | Facts pulled from somewhere else |
| Fixed text | Each claim, quote, number and link that must stay word for word | A claim drifting from "active Content ID claim" to "copyright claim" |
| Permitted changes | What may change (order, line breaks, opening line) and what may not (new numbers, quotes, examples, hashtags, emojis) | Additions that sound plausible but have no source |
| Audience and voice | Who the post is for and three or four words on how you sound | A generic engagement voice that is not yours |
| Missing-evidence rule | If the post needs a fact the source lacks, leave it out and add a line starting MISSING: | An invented statistic to satisfy the request |
The prompt used in the test
This is the full controlled prompt from the test below, with the source post in place. Replace the source, the fixed text, the audience and the length for your own post. Keep the rules.
- Rewrite the source post below for LinkedIn.
- Source post (the only source of facts): "Music rights don't travel with the clip. YouTube says Shorts longer than one minute with an active Content ID claim are blocked worldwide. Check the rights for each place you repost before you export. ferryman.io/blog/repurpose-one-idea-across-social-platforms"
- Facts that must stay exactly as written: "YouTube says Shorts longer than one minute with an active Content ID claim are blocked worldwide." and the link ferryman.io/blog/repurpose-one-idea-across-social-platforms
- You may change: sentence order, line breaks, the opening line, and add one sentence of context for people who manage video for a brand. You may not: add numbers, statistics, quotes, examples, hashtags, emojis or claims that are not in the source. Do not say what YouTube does beyond the quoted rule.
- Audience: marketing managers who repost short videos across platforms. Voice: plain, direct, second person, no hype words. Length: under 600 characters.
- If the post would need a fact the source does not contain, do not invent it. Write the post without it, then add a line after the post starting with "MISSING:" naming what is missing. Output only the post, followed by any MISSING line.
A real test: one post, two prompts, three runs each
The source was one of Ferryman's own scheduled brand posts. On 2026-11-16 we sent it to Claude Sonnet 4.5 (claude-sonnet-4-5-20250929, default settings, no system prompt) three times with a one-line prompt, "Rewrite this post for LinkedIn so it gets more engagement", and three times with the controlled prompt above. One model, one post and six runs: this shows what to look for, not how often any model fails. Outputs vary from run to run, so your results will differ.
| Run | Prompt | YouTube sentence kept exactly | Link kept | Added without a source | Ready to post as returned |
|---|---|---|---|---|---|
| 1 | One line | No: "active" dropped, "one minute" became "60 seconds" | Yes | "I just learned", a claim about TikTok, emojis, 5 hashtags | No: preamble and a "why this works" list |
| 2 | One line | No: "active" dropped, split across bullets | Yes | "I just learned", "might not be legal everywhere", a "copyright strike", emojis | No: preamble and a "why this works" list |
| 3 | One line | No: "YouTube now blocks", "new policy", "copyright claims" | No: replaced by [link] | A recent policy change, "Just lost a high-performing video?", emojis, hashtags | No: three options instead of one post |
| 4 | Controlled | Yes | Yes | One generalizing sentence: "What works on one platform may not be cleared for another" | Yes, 408 characters |
| 5 | Controlled | Yes | Yes | The same kind of generalizing sentence | Yes, 444 characters |
| 6 | Controlled | Yes | Yes | Nothing | Yes, 361 characters |
What went wrong with the one-line prompt
None of the one-line outputs was wrong about the main idea, which makes the errors easy to miss. Each one changed a detail. Dropping "active" widens the rule to any Shorts that ever had a claim. "Copyright strike" names a different YouTube process from a Content ID claim. "YouTube now blocks" and "new policy" assert a recent change the source never mentions. Two runs invented a first-person story, "I just learned", in the author's voice.
The engagement request also pulled in a house style: warning-sign emojis, bolded hooks, a closing question and hashtags. That is a voice, just not the author's. The third run dropped the link entirely and offered three versions, so someone in a hurry could publish a post with "[link]" in it.
The two generalizing sentences in the controlled runs are a judgment call. They restate the source's advice to check rights per destination rather than add a new fact, so they pass the rubric below. A stricter prompt would forbid any added sentence.
Testing the missing-evidence rule
We then added one line to the controlled prompt asking the model to open with a statistic on how many creators the rule affects. The source contains no such number. In all three runs the model wrote the post without a statistic and named the gap on a MISSING line. In one run the flag came first, in bold, above the post instead of after it, so check the whole output for the flag, not only the last line.
This is the most useful part of the prompt. A request you can't satisfy from the source should come back as a question for you, not as a confident number.
A prompt keeps facts; it does not check them
While auditing the test we re-read the YouTube Help page behind the source post. On 2026-11-16 one of its answers says: "Shorts longer than one minute that have an active Content ID claim, regardless of the policy, will be blocked on YouTube." Our first re-read stopped there, so we took "worldwide" for an error and reworded the scheduled post and the guide it links to. A full read of the same page later that day found the other sentence: a Short over one minute "with an active copyright claim of any type, including manual claims, will be blocked globally on YouTube." The original wording had a source after all. The reworded versions quote the page too, so they stayed.
So the order matters. Check the source against its own source first, reading the whole page, then ask the model to preserve it. A prompt that preserves facts makes a model faithful to your draft, including its mistakes, and a correction needs the same check as the claim it replaces.
Review rubric for every AI rewrite
Read the output next to the source. One failed check means edit or regenerate; don't publish and fix later.
| Check | How to test it | Fail looks like |
|---|---|---|
| Fixed text intact | Find each fixed sentence, quote, number and link in the output, character for character | A word dropped, a number converted, a link replaced by a placeholder |
| Nothing new | Underline every claim in the output and find it in the source | A statistic, a policy change, a quote or an anecdote that is not in the source |
| Still your voice | Read it aloud. Would you say this sentence? | "I just learned" when you didn't, hype words, emojis you never use |
| Flags handled | Search the output for MISSING anywhere | A flag you didn't read, or a gap filled with a guess |
| Fits the destination | Check the length limit and format for where it is going | Over the character limit, markdown that will show as asterisks, extra options or commentary |
| Source still current | Re-open the page behind each factual claim before publishing | The source changed after you wrote the original |
Where Ferryman fits
In Ferryman's sync routes you can give each destination its own AI instruction of up to 500 characters, such as "Make this more formal for LinkedIn". Ferryman runs it with Claude Sonnet 4.5 and fixed rules: rewrite only the supplied post, keep factual claims, links, @mentions and #hashtags, don't invent facts, metrics, quotes, links or calls to action, and stay within the destination's character limit, unless the route splits long posts into threads. Your original post is not changed.
Ferryman also checks each rewrite in code. If it drops a link that starts with http:// or https://, invents a new link, drops an @mention or #hashtag, comes back empty or with a preamble such as "Here is", or runs past the destination's limit (ten posts' worth when thread splitting is on), the destination receives your original text instead. Those checks catch structural mistakes. They can't tell whether a reworded sentence still means the same thing, so the rubric above still applies to what you publish.
FROM GUIDE TO WORKFLOW
Explore this workflow in Ferryman
Give each destination its own AI instruction
Write a short instruction per destination for tone, length or structure. Ferryman keeps your original post unchanged and sends the original instead when a rewrite drops a link, mention or hashtag or runs past the platform's limit.
Sources & further reading
Written by Ferryman. Sources checked Nov 16, 2026. Platform rules can change; consult the linked documentation for current requirements.