Skip Telegram Photos With a Word Printed in the Image

To skip a Telegram photo post because a word such as PROMOTION is printed inside the picture, give a text rule something to read. In RedFox Auto Forward Telegram, an OCR Image AI Custom Prompt returns a unique marker (for example OCR_SKIP_PROMOTION) only when the image contains the word, APPEND TO CAPTION writes that marker into the message content, and a Basic Blacklist rule on the same task matches the marker. RedFox documents this chain in its image-keyword walkthrough. A marker is better than blacklisting the ordinary word: a plain rule would also block captions that merely mention it, and a caption rule cannot read pixels at all.

This guide is for an existing RedFox Auto Forward Telegram route that already delivers photos, where the unwanted wording sits in the image and not in the caption. The chain has three separate failure points—the model’s reading, the appended text and the Blacklist match—and a missing post cannot tell you which one failed. Below is a private test that exercises each link on its own, with a decoy for every shortcut. It is not a general content-moderation system.

Identity and scope: use @AutoForwardNew_Bot. The previous @Auto_Forward_Messages_Bot has moved to this current handle. RedFox develops the product; Auto-Bot.io publishes this editorial guide and does not operate the bot. Sources were checked on October 28, 2026. The specimens and expected results are a proposed reader exercise, not an executed test, and the featured artwork is a concept illustration, not product UI.

1. Decide what the rule can see: the picture, the caption or neither

A word printed in a photo is not a caption. Our photo and caption diagnostic, citing Telegram’s Message documentation, treats a photo and its caption as separate fields. That is why the obvious shortcuts fail for this reader:

• Blacklist the word promotion. It can see message text and captions. RedFox’s Blacklist guide shows promotion,unsubscribe blocking a message that says “Check out our latest promotion!” That would block a caption that mentions the word, yet nothing reads the pixels of a photo with a neutral caption.
• Exclude photo types. It sees the kind of message, not its content. As that photo guide summarizes RedFox’s type filters, excluding photos removes every photo and Photo Only removes every captionless one. Both are blunt: a harmless photo goes too.
• OCR marker plus Blacklist. It sees the text a model returns from the picture once it is appended to the message. It filters on the image-derived signal alone, but only as well as the model reads and only if the marker really reaches the rule.

If the wording you care about is in the caption, use a text rule instead; whole-word rules cover precision there. If you want a photo’s printed facts shown as readable text, that is a different job, covered by the event-poster OCR guide, which prepends extracted text rather than filtering on it.

2. Check the option and your prerequisites first

RedFox’s own pages disagree, and which one is right for your version decides whether this guide applies. The general OCR Image AI page documents Prepend to Caption as the output position and lists Append to Caption as a possible future option. The keyword walkthrough instead tells you to select APPEND TO CAPTION under Text Output Position and describes the marker filter built on it. The saved copies of both pages show no revision date, so treat them as undated snapshots from the day this was written.

Treat the walkthrough as the documented path only if your task’s Text Output Position actually offers APPEND TO CAPTION. If it offers only Prepend, stop and ask RedFox support to confirm the option for your app version. Neither page says that Blacklist inspects prepended text, so a Prepend variant would be an undocumented experiment and you would need to run the whole private test below to learn anything from it. Do not assume either page is the newer one.

Also confirm, before you spend time:

• Access and billing. RedFox’s README lists AI Mode under Platinum, and the OCR page asks for an OpenAI key with vision access. RedFox’s API key guide says the key needs billing and that reading images uses more tokens than plain text. Check your own plan and key rather than assuming either.
• Reach. OCR Image AI is switched on per task, and the OCR page says it processes incoming images through the selected model. The pages read do not describe limiting it to particular posts, so every image on that task is in scope for cost, delay and what you are willing to send to the model.
• Blacklist reaches the task. The task’s Advanced Configuration has a Black List option described as allowing the Blacklist feature to work on that task. A Blacklist rule is also a reusable object with Activate/Deactivate For Task and Apply For All Tasks controls, so it can be live on more tasks than the one you are editing.
• Management surface. Prefer the Web app, iOS app or Android app for fuller management than the limited Telegram-bot entry point. Active tasks run in RedFox’s cloud; the runtime FAQ says your own internet does not have to stay on.

3. The documented chain, with the marker kept unique

The walkthrough’s prompt is below. It asks the model to read the image text and return the marker only when the word is present; otherwise it asks for an empty response.

Read the text in the image.

If the image contains the word "PROMOTION", regardless of capitalization, return exactly:

OCR_SKIP_PROMOTION

If the image does not contain the word "PROMOTION", return an empty response.

Return only the required result. Do not provide an explanation or any other text.

The setup is then three saved settings on one task: OCR Image AI enabled with Custom Prompt and this text in Prompt Request; Text Output Position set to APPEND TO CAPTION, saved with Save & Apply; and a Basic Blacklist rule whose content is the exact marker. RedFox also advises a marker that is unlikely to appear in normal captions, and warns not to add generic words such as OCR_PASS to Blacklist.

Two things the prompt leaves for you to decide. First, “contains the word PROMOTION” does not say whether PROMOTIONS or PROMOTIONAL count; write your policy down and test those spellings rather than letting the model decide. Second, the marker is plain text in the message: anyone who types it in a caption triggers the rule, with no OCR involved. That collision can affect any source whose caption happens to contain the marker. It is also a convenient way to test the rule separately, as the specimens below do.

4. Test each link on a private task

Use a source and a destination you control, with no audience, and images you made yourself. Do not start on the live feed: Stage 1 below deliberately delivers captions that carry the marker. Before preparing the baseline, record the task’s label, source, target, Black List gate, rule associations, type filters, Cleaner, Replace, other transformations, Duplicate Filters, delays and schedules. Record the selected model, prompt and output position too. Prepare this isolated private task with Black List already on, no marker rule associated and all other associations fixed at their intended test values. Keep Duplicate Filters off on this isolated experiment so fresh copies of the same specimens remain eligible; never loosen a live task to do this. Finish any preparatory changes before establishing Stage 0, then hold these settings, sender and rule matches stable between stages.

The specimens

• T — route control. A plain text post, caption “ROUTE CHECK T”. If T fails, no later result means anything.
• A — positive. A large, sharp image printing “PROMOTION”, caption “Weekly update”. The word is only in the picture.
• A2 — captionless positive. The same image as A with no caption. The sources do not describe appending to an empty caption, so do not assume it.
• B — negative control. The same layout printing “SCHEDULE”, caption “Weekly update”. It must keep arriving, clean.
• C — caption-only decoy. The same “SCHEDULE” layout with the caption “Promotion ends Friday”. The word in the caption must not be the trigger.
• D — typed-marker decoy. The same “SCHEDULE” layout with the caption “Weekly update OCR_SKIP_PROMOTION”. It tests the Blacklist rule with no OCR involved and shows marker collision.
• E — hard case. A deliberately blurred, rotated or partly covered “PROMOTION”, caption “Weekly update”. RedFox says to use larger, sharper, high-contrast text and to avoid blurred, rotated or partly hidden words.
• F — spelling variants. “Promotion”, “PROMOTIONS” and “PROMOTIONAL” on separate images, caption “Weekly update”. Set the intended outcome for each spelling first, make the prompt explicit if necessary, then use these specimens to check that policy; the examples do not encode it by themselves.

Stage 0: establish the prepared eligible baseline

With the prepared Black List gate already on, OCR off, no marker rule associated and the recorded settings fixed, send fresh copies of every specimen. All of them should arrive. This confirms that each one is eligible for the route; a specimen that never arrives at baseline cannot later be called “filtered”. Check the failures against your existing rules (for example a caption rule already containing promotion) or use the photo and caption diagnostic before continuing.

Stage 1: observe the OCR output, with no Blacklist rule yet

Enable OCR Image AI, choose Custom Prompt, paste the prompt, set APPEND TO CAPTION and Save & Apply. Reopen the saved model, prompt, enable switch and position and confirm they match before sending fresh copies. Because APPEND TO CAPTION is documented as adding the extracted text to the end of the forwarded message content, the expected shape is that A arrives with OCR_SKIP_PROMOTION added, while B and C arrive without it. Your destination message is the authority on separators and layout. Allow time for processing; RedFox notes that AI may delay forwarding.

This stage gives a clean initial observation of the appended output before the marker rule is associated. A later rule-off reading shows only that later run’s output, not what the model returned on an earlier missing run. A seen in the destination with the marker is evidence that, for this specimen, the model read the word and the append step worked. It is not yet evidence that anything is filtered.

Stage 2: add the rule, one change

Create a new Basic Blacklist rule with its own label, the exact marker as its content, Apply To set to Post, and association with this task only. Use a new rule and leave existing rules and Apply For All Tasks alone, because editing a reusable rule can reach other tasks. Leave the general Black List gate at its baseline on value: associating this new rule is the only change in this stage. Send fresh copies of every specimen. The expected result is that A and D do not arrive, and T, B and C still do. B, C and T are your retained delivered controls: they show that the rule is not simply removing everything.

Reading the results

• Stage 1: A arrives with the marker; B, C and T are clean. The model returned the marker and the append step worked for A. Proceed to Stage 2.
• Stage 1: A arrives without the marker. The new marker rule cannot be responsible because it is not associated yet. Check OCR enable/save state, prompt, position, model and processing delay, along with any recorded transformation that could remove output. Change one suspected setting, resend a fresh eligible A, and restore an unexplained adjustment before another trial.
• Stage 1: B or C arrives with the marker or other text. The model is returning text when the word is absent, and that text would reach your audience. Tighten the prompt or try another image-capable model; do not attach the rule yet.
• Stage 2: A and D are missing; B, C and T arrive. Together with Stage 1, this is consistent with marker-based blocking, not direct observation of an internal match event. Complete the rule-off/on reversal below and check E and F against your written policy.
• Stage 2: A and D both arrive. If D retains its exact typed marker, the rule did not block that specimen as intended. Check task association, the unchanged Black List gate, Post scope and spelling; inspect A’s current output separately. Arrival alone does not identify the failed stage.
• Stage 2: D is missing but A arrives with the marker. This suggests a difference between the two specimens, but D’s absence alone does not expose a match event. Possible causes include a different string (spacing, case), another exclusion or the appended text not reaching the rule. The walkthrough describes the intended chain but does not specify a general processing-order contract. Compare the delivered marker character by character, then send both specimens to RedFox support. Change nothing else.
• Stage 2: B or C is missing. Investigate a suspected false skip or a delivery failure. Detach only this task’s marker rule and resend a comparable fresh specimen to inspect the current appended output. A later reading does not reveal what the model returned on the earlier missing run.
• T is missing at any stage. Invalidate that stage. Check saved task state, endpoints, sender/message eligibility, rules, delays and visible errors using RedFox’s not-forwarding checklist. This control failure neither identifies nor rules out OCR or a rule.

Confirm the rule with an off/on reversal

Keep OCR and every baseline setting unchanged. Deactivate only the marker-rule association on this private task. Send fresh comparable A and D plus T, B and C. A should return with the exact appended marker and D with its typed marker; T, B and C should still arrive clean. Restore that association, wait for processing to settle and repeat with fresh copies: A and D should again be absent while the controls arrive. Record times and visible errors. This reversible pattern supports a scoped filtering decision, not a guaranteed internal trace or universal accuracy. If it does not hold, restore the baseline and leave the cause unresolved; do not recommend a live change.

Stage 3: rehearse OCR-off, then rule-off

RedFox treats the two controls separately: turning OCR off does not remove the Blacklist entry, and removing the entry does not turn OCR off. With the marker rule still associated, turn OCR off, save and reopen to confirm. Allow earlier processing to settle and account for the specimens already sent; the sources do not say that disabling OCR cancels queued or in-flight work. Send fresh eligible A, B, C and T: they should arrive without added text. D still contains the typed marker and should remain excluded. Only then deactivate the marker rule for this task and repeat the controls, including a fresh D, which should now arrive with its intentionally typed marker. Do not delete a rule used elsewhere. If prior work cannot be accounted for, keep the protective rule and seek support before a live transition.

5. Decide what to do with misses and false skips

E and F are not pass/fail by default; they tell you what the model does with your real artwork. OpenAI’s image-input guide lists small text, rotated text, non-Latin alphabets such as Japanese and Korean, and incorrect descriptions among the limitations of its vision models. RedFox’s OCR page asks for an OpenAI key, but your task’s chosen model is what you are testing. Run the specimens in the language and style of your real feed.

The two errors cost differently. A missed word delivers the post you wanted gone. A false marker removes a post you wanted to keep, and the sources describe a skipped post as one that does not arrive; the pages read do not describe a skipped-post log or review queue. So for content where a wrong skip or a missed one matters (legal, safety, regulated or contractual material), do not let this filter be the only control. Spot-check the source against the destination on a schedule you define, and keep a human review outside the product for ambiguous or unreadable images.

An independent open-source project shows the same concern from another direction. The Telegram-OCR README describes a Tesseract scanner that exports Telegram messages whose image text matches your keywords, notes that matching tolerates OCR errors, and exposes a match-sensitivity threshold. It is a different tool with a different job, it needs your own Telegram API credentials and a login code, and we have not tested it. We cite it only as evidence that recognition error is treated as a normal part of image-keyword work, not as a recommendation or a measure of demand.

6. Roll out and roll back in order

After the private run, apply only what the evidence supports to the intended live task. Use a new rule associated with that task only, keep the previous values for rollback, and confirm that no other task already uses the marker. Then inspect one fresh approved image post that should be skipped and one that should be forwarded.

Our inferred change order reduces marker exposure; it is not a RedFox guarantee or an atomic update. In a quiet controlled change window, associate the validated rule first, then enable OCR and confirm the saved state. To undo it, turn OCR off first and retain the rule until earlier queued or in-flight processing has settled or been accounted for, then detach it. Stage 3 rehearses that order privately. Do not assume a save cancels earlier work or promise a clean transition if you cannot establish what remains in flight.

Several cases are not covered by the sources. RedFox’s Blacklist guide describes all-or-nothing albums in its video/document-media-filter section; that does not establish how a Basic OCR-marker rule handles albums or whether OCR runs on each member. Test albums separately. Images sent as files and Read Text in Stickers with AI are separate matters, and the sticker option is its own switch on the OCR page. The pages describe testing with new posts; nothing read says earlier forwarded posts are scanned or removed. To change the keyword and marker, use the same controlled OCR-off transition, account for earlier work, prepare the new prompt and task-scoped rule, and validate fresh specimens before enabling it for a live feed. Two separate saves are not a simultaneous change.

FAQ

Why not put PROMOTION itself in Blacklist?

Because it matches message text, and the word you are filtering is in a picture. If OCR text were appended and you blacklisted the plain word, a caption such as “promotion ends Friday” would also be blocked. The marker keeps the rule tied to the image-derived signal. Specimen C checks that.

Why does the guide insist on APPEND TO CAPTION?

Because the walkthrough is built on it, while the general OCR page lists it only as a possible future option. If your task does not offer it, the documented path is unavailable in your version, and a substitute is an experiment to prove with Stages 1 and 2.

What if non-matching images get text appended anyway?

Tighten the prompt so it demands an empty response, or test another image-capable model. RedFox also mentions a non-blacklisted placeholder such as OCR_PASS for a model that cannot return an empty response, but warns it may be appended to the forwarded caption. That puts text on every ordinary post, so weigh it against a model change.

Can the keyword be a phrase, or several words?

RedFox says a phrase works by replacing the word in the prompt and updating the marker. Several unrelated words in one prompt are not covered, so treat that as a new design: one marker per decision and a full run of the specimens.

Prove the chain on one task before it filters a feed

Open the RedFox Auto Forward Telegram overview, then use the Web app or your preferred mobile app to check whether your task offers APPEND TO CAPTION. Your useful result is specific: an eligible baseline, A arriving with the marker without the rule, a successful rule-off/on reversal with A and D excluded only while the marker rule is associated and T, B and C still arriving, a policy checked against blurred and variant spellings, and an OCR-off/rule-off rollback rehearsed with earlier work accounted for.