How-To Guide

How should I give feedback when an AI social post is technically correct but off brand?

Turn 'this doesn't feel like us' into a four-field revision brief with an observable acceptance test, bound to a version.

9 min read

By the Superdirector editorial team. The feedback, brief and approval pages cited here were read November 15, 2026, as were Meta's developer documentation and the Superdirector pricing page. Every worked example is a labelled hypothetical, not a customer account.

At a glance

How should I give feedback when an AI social post is technically correct but off brand?

Treat the correction as a revision brief, not a complaint. Name the written brand rule the post broke, quote the exact line that broke it, give one on-brand replacement or a short on-brand and off-brand pair, and state the observable test the corrected version must pass. Keep the brief to four fields on one screen, bind it to the version you read, and check the changed lines against those four fields on the next version instead of rereading the whole draft. A brief that says what to change, why that rule applies, and how you will know it is fixed is actionable by a human writer or an AI operator. A brief that says 'make it better' hands the work back as a guess.

  • Separate a factual error from a voice miss: a post can be accurate, on budget and still wrong for the brand.
  • Use four fields: the rule, the exact line, one example, and the acceptance test.
  • Make the test observable, something two reviewers can settle without arguing about taste.
  • Show an on-brand and off-brand pair instead of an adjective; tone descriptions drift, examples converge.
  • Bind feedback to a version and recheck only the changed sentences against the four fields.
  • In an AI-run account the same brief works; Superdirector (ours) publishes eligible Instagram carousels and single X posts under the account's standing permission, not a per-post approval queue.
A four-row revision brief card labelled rule, line, example and test, beside a struck-through note reading this does not feel like us.

Four labelled fields turn a vague complaint into an actionable brief. Original diagram, drawn from this guide's own four-field brief.

Superdirector

Make it work for your brand

Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.

Separate a voice miss from a factual error

The post is accurate. The product name is right, the price matches the source, the link opens the right page. It still does not sound like the brand. That is the gap this guide is about, and it is the gap most feedback fails to close, because the reply reads 'this does not feel like us' and the next draft is a different guess. The Post Engine's quality checklist names the exact case: "Even when everything is technically correct, the post may still sound like it was written by a machine."

Before writing the feedback, sort the flaw into one of four buckets. A wrong fact, a promise the business cannot support, a product detail that misstates what ships, or a voice that belongs to a generic landing page. The first three are correction decisions with their own rules and their own urgency; this guide covers the fourth. Sorting first matters because a voice rewrite must not quietly change a fact, and a factual correction should not be filed as a tone preference.

  1. Fact: a number, date, price or name that does not match its source. Fix it through a correction rule, not a style note.
  2. Promise: a claim about results, guarantees or performance that the evidence does not support. Treat it as a claim decision.
  3. Product: a feature, colour, size or compatibility detail that is not what ships. Check it against the product source.
  4. Voice: the facts hold, but the wording, rhythm or framing does not match the brand on that channel. This is the feedback brief.

Air's analysis explains why the voice bucket fills up so often: "AI-generated content doesn't match your brand because the model has no working memory of it, not because someone wrote a short, thoughtless prompt." The model is not ignoring you. It never received the rule, the example or the channel context in the first place, so the correction has to carry them.

Write one revision brief with four fields

For a single post you do not need a long document. You need four fields that fit on one screen and survive being forwarded to a writer, a freelancer or an AI operator. Niko Alho's brief template lists eight fields for a full article and places acceptance criteria last: "A working brief has 8 fields: intent, target query, audience operator role, entities to cover, evidence requirements, structural skeleton, tone exemplars, and acceptance criteria." For a correction, four of those carry the weight.

  1. Rule: the written brand rule the line broke, for example no superlatives the product cannot prove, or openings must name the reader's situation.
  2. Line: the exact sentence or fragment, quoted, so nobody has to guess which part is wrong.
  3. Example: one on-brand replacement, or a short on-brand and off-brand pair when the problem is taste, not a banned word.
  4. Test: the observable acceptance criterion the corrected version must pass before it can publish.

Ziflow's guide is blunt about why the rule field cannot be skipped: it tells teams to "Establish clear evaluation criteria", then to check whether the post uses "messaging, vocabulary, and product positioning consistent with their brand?" Its warning that AI creative is "notorious for missing the mark here" is the reason a reviewer writes the rule down instead of assuming the writer knows it.

The table below is a labelled hypothetical, not a real client or a real post. It shows the same complaint in its unusable form and in the four-field form.

FieldUnusable noteRevision brief
RuleMake it more exciting.Openings name the reader's situation; no superlatives the product cannot prove.
LineThe whole thing feels flat."Our revolutionary brew will transform your mornings forever."
ExampleYou know what I mean."Your 6am kettle deserves better than a pod." Keep the plain, specific voice of the approved May launch post.
TestMake it sound like us.No 'revolutionary' or 'forever'; the first sentence names a morning situation; the post reads like the approved May launch exemplar.

Note what the right-hand column does that the left cannot: it names a rule an editor can cite, a line a writer can find, an example a model can imitate, and a test that ends the argument. Predis's approval guidance adds the record habit that keeps this auditable: "keep detailed documentation on who requested the change, why, and how it was implemented." The four fields are that documentation in miniature.

Make the acceptance criterion observable

An acceptance criterion is observable when two reviewers can reach the same verdict without negotiating taste. 'Sounds more like us' is not observable, because it moves with whoever reads it. Niko Alho states the failure mode and the fix in one line: "Tone descriptions get interpreted differently every run." The answer is not a longer adjective list; it is a test with inputs and a verdict.

Air's pre-publish checklist supplies two questions worth turning into tests: "Does the copy sound like the brand on this channel?" and "Would someone on the team recognize this as your brand with the logo removed?" Both are judgments, but both can be recorded as a named reviewer's yes or no, which already beats a shrug. The Post Engine adds the reason to keep the language specific: "Specific language makes the content easier to understand and less interchangeable."

Write the test as a short ordered list, and make each line something a person or a script can run:

  1. Banned phrases: scan for the words the brand has retired, such as 'game-changing' or 'in today's fast-paced world', and fail on any hit.
  2. Required vocabulary: the product, the audience and the category are named the way the brand names them, with no invented synonyms.
  3. Opening pattern: the first sentence does the job the brand's openings do, for example naming the reader's situation, not the product.
  4. Length and rhythm: the post sits inside the channel's length range and avoids the long unbroken sentence pattern the brand has flagged before.
  5. Exemplar match: a named reviewer compares the draft with one approved post and records pass or fail, with the failing line quoted.

One caution belongs with the test: an AI can help spot a claim that needs checking, but it should not be the only authority that signs it off. The Post Engine puts it directly: "AI can help identify claims that need verification, but it should not be treated as the only authority". The same limit applies to voice. The model can imitate an exemplar; a person decides whether the imitation is good enough for the account.

Show a pair instead of an adjective

The fastest way to make a voice rule concrete is to show two versions of the same post. Julie Holmes calls feedback calibration, and her example of a usable correction is a rewrite instruction, not a verdict: "Instead of 'make this better,' say 'add real-world examples and keep it conversational.'" The contrast does the teaching that an adjective cannot.

Here is a labelled hypothetical for a small coffee subscription. Both lines are invented for illustration, and neither is a customer post. The off-brand line is not factually wrong; it is the voice of every other subscription ad.

VersionLineWhy it passes or fails
Off brandUnlock the ultimate coffee experience and transform your mornings with our revolutionary subscription.Three superlatives, no specific product, and a promise the product cannot prove.
On brandYour 6am kettle deserves better than a pod. Fresh-roasted beans, ground for your brewer, at your door before the first cup.Names a real situation, a concrete product detail and a benefit the service actually delivers.

Holmes's wider point is that this kind of correction is not a final polish. Her 20-60-20 model gives review and refinement the last 20 percent of the work, and says that share is where feedback lives. A pair of examples is the cheapest way to spend that 20 percent well, because it gives the writer something to copy rather than a mood to guess.

Keep the pair attached to the rule. If the same brand keeps rejecting the same kind of line, the examples belong in stored brand context, not only in this week's message. The brand voice guide and the training workflow are where that durable copy lives; this note is the correction that sends the writer back to it.

Two caption cards side by side, an off-brand line reading ultimate coffee experience against an on-brand line reading your 6am kettle.

The off-brand line leans on superlatives; the on-brand line names a morning situation and a concrete detail. Original diagram, drawn from the guide's labelled coffee-subscription pair.

Bind the feedback to a version and recheck only what changed

Feedback without a version is a memory test. Predis's video approval guide begins with the control that prevents drift: "Have a version control policy", and its documentation rule says to "keep detailed documentation on who requested the change, why, and how it was implemented." That record is what lets you say which version the brief applied to, and which version the corrected draft is.

Ziflow names the practical version of the same habit: "Track changes, view revisions, and compare versions." When the corrected draft arrives, do not reread it from the top and hope the feeling has changed. Open the brief, find the changed lines, and mark each of the four fields kept, still failing or unresolved. That is a short, decisive read.

Two process rules keep the loop from stalling. First, a correction that fails the same rule twice goes back to the brief, not to the writer: Niko Alho's rule is that when the writer is going off course, the fix is upstream. Second, keep a human override and an escalation path, as Blaze's workflow guidance recommends when it says to "Design workflows with 'human-in-the-loop' checkpoints" and to "Allow for manual overrides or escalation paths". A post that still reads wrong after two focused rounds is a decision for a person, not another automated pass.

If you need the next step after the correction, that is the reconciliation read: checking whether the operator actually applied each finding to the new version. It is a separate check from composing the brief, and it is covered in the linked guide on telling whether an AI operator applied your feedback.

Four steps in order, reviewed version, four-field brief, corrected version and acceptance test, with a branch back to the brief when a rule fails twice.

Feedback stays bound to a version, and a rule that fails twice sends the work back to the brief. Original diagram, drawn from this guide's version-bound process.

Where Superdirector fits and where it does not

Superdirector (ours) sits on both sides of this loop without absorbing the feedback decision. It produces the drafts you would correct: it watches what is working in your niche, makes branded carousels and X posts, and fills your calendar. On the publishing side, the pricing page is specific: Growth at $199 a month is one AI-run operated account, and extra accounts are $99 a month each. It "publishes eligible Instagram carousels and single X text posts by themselves once posting is turned on for an account." Starter at $29/month is the hands-on creation tier that works from Claude or ChatGPT, not the operated publishing account.

That publishing is account-level. The same page says "Posting starts only when you turn it on for an account, after one warning" and "You make the important calls: what posts, what changes and what gets paid." There is no per-post approval step inside the publishing path, so the feedback brief stays with the people running the brand. It is how you correct a draft before it goes out, and it is how you judge one that did.

What Superdirector does not do is decide that a line is off brand for you. It can hold the brand context and the exemplars, and on Growth it can publish once you have turned posting on for the account, but the rule, the example and the acceptance test remain the brand's. The pricing page also asks you to "Label AI content where a platform requires it", which is a pre-publication decision, not a cleanup.

The honest way to use this guide: if your recurring problem is a stream of posts that are accurate but generic, buy the four-field brief habit first, because it is free and it works with any writer or tool. If your problem is an empty calendar and you want the drafts and the publishing handled under an account permission you set, Superdirector is built for that, and the same brief still tells it what 'on brand' means when something has to be corrected.

Frequently asked questions

How should I give feedback when an AI social post is technically correct but off brand?

Send a revision brief with four fields: the written brand rule the post broke, the exact line that broke it, one on-brand example, and an observable test the next version must pass. Bind it to the version you reviewed, and check only the changed lines when the correction comes back. That structure gives a writer or an AI operator something to act on and gives you something to verify. 'This doesn't feel like us' does neither.

What should a revision brief for an off-brand post contain?

At minimum: the rule (for example, no superlatives the product cannot prove), the quoted line, one replacement or a short pair of examples, and the acceptance test. Niko Alho's content brief lists acceptance criteria as one of its eight fields and says: 'If acceptance criteria are not in the brief, you have a creative writing prompt, not a content pipeline.' For one post you do not need all eight fields, but you do need those four.

How many examples should I give?

One good replacement plus, when the failure is a matter of taste, one contrasting bad line. Two short examples are usually enough because the contrast carries the rule. Julie Holmes's coaching advice is to replace 'make this better' with a specific instruction such as 'add real-world examples and keep it conversational.' If you cannot produce an on-brand example yourself, the brief is not ready, because the standard is not yet written down.

How do I make 'sounds like us' testable?

Convert it into checks: a banned-phrase scan, required vocabulary, an opening pattern, a length range, and a comparison against one approved exemplar. Air's pre-publish checklist uses a strong recognition test, whether someone on the team would recognize the brand with the logo removed, and asks whether the copy sounds like the brand on that channel. Those questions still need a named owner and a recorded verdict to become an acceptance test.

What if the AI keeps missing the same rule?

Fix the input before you blame the model. Niko Alho puts the cause plainly: if the writer is hallucinating or going off-topic, the fix is upstream, and tone descriptions get interpreted differently every run. Add a sharper exemplar, move the rule into stored brand context, and rerun one draft against the same test. Two failures of the same rule in a row mean the brief is underspecified, not that the writer is hopeless.

Should feedback arrive as one message or line by line?

One brief with numbered findings, not a chat thread. Ziflow's feedback workflow separates evaluation criteria, a first review round, a cross-functional round and revision, then asks the platform to track changes and compare versions. A single numbered brief maps cleanly onto that record, and each finding can be marked kept, changed or unresolved on the next version.

Does Superdirector hold each post for approval before publishing?

No. Growth at $199 a month is one AI-run operated account, and extra accounts are $99 a month each. It publishes eligible Instagram carousels and single X text posts by themselves once posting is turned on for an account, after one warning that names the account, the public effect and the cost. There is no per-post approval queue in the publishing path. The control is account-level, and the pricing page keeps the standing calls with you: what posts, what changes and what gets paid. Starter at $29/month is the hands-on creation and MCP tier, not the operated account.

How is feedback to an AI operator different from feedback to a human writer?

The brief is the same; the delivery and the memory are different. A person remembers last month's correction, while a model starts from the same general average on the next run unless the rule and examples live in stored brand context. Air puts the cause at the model's lack of working brand memory, and The Post Engine notes that AI does not always signal its own uncertainty. So write the rule down once, attach the example, and reuse both on the next post instead of relying on memory.

Related practical guides

Check that the operator applied the feedback

Turn the four-field brief into one numbered finding per field, pin the draft version you reviewed, and ask for a reply that answers each finding with its status and its changed sentence. That reconciliation read is separate from writing the brief.