Learn
Learn: how we make ads with AI, from our own production
Guides on making video ads with AI, written from the ads we make ourselves: the workflow, the models, the formats, your own footage and the tests after launch. Each one says what we do, where it broke for us and what we changed.
Early access we set it up with you. Coming we’re building it.
We are the team behind Turntake. Turntake started as the machine behind the ads we make for SmartSolar, in Denmark, Sweden and Ireland: talking heads, claymation, skeletons, crochet, 3D animation and drone explainers, in three languages. The machine behind it has been running since June 2026, and Turntake grew out of it when editing became the bottleneck.
These guides come from the same production. They describe what we do and the rules we keep, including the places where it broke for us: a model that refused our faces, a voice that needed numbers written out, captions that had to be retimed to the voice. When a guide describes a model, the facts come from the maker’s own documentation, linked in the text. When it describes our own work, it says so, with a date.
The ad factory
The whole chain, from research and script to voice, talking head, b-roll, edit, checks and test. Start with the A to Z workflow, then go deeper into the parts: notes on a video, the checks before upload, hooks, and the formats we make, from talking heads to clay and skeletons.
- Start here · 7 min readHow to make AI UGC ads: the full workflow, A to ZOur A–Z workflow for AI UGC ads, from research and script to voice, talking head, b-roll, edit, QA and test. What each step needs and where it breaks.
- Guide · 5 min readHow to give feedback on a video ad: notes that workPin each note to a second, a word, a clip or a spot in the frame, send them as one batch, and check the next take against the list. Notes that get used.
- Guide · 6 min readReels safe zones and our video ad QA before uploadKeep captions, faces and buttons clear of the Reels overlays, then check loudness, caption lines, frozen frames and shot length. Our house rules.
- Guide · 6 min readAI UGC talking heads: a still, a voice and b-rollHow we make AI UGC talking heads: an approved portrait plus your own voice track, b-roll from a video model, and the tells we cut so it doesn’t feel off.
- Guide · 5 min readAI skeleton ads: how to make the format work as an adAI skeleton ads from our own production: a locked character, one place per shot, one continuous voiceover, no numbers in the frame and a reusable library.
- Guide · 5 min readAI claymation ads: build the clay world once, reuse itMake AI claymation ads as a world you build once: one character and style block in every prompt, silent figures, a narrator and a clip library you reuse.
- Guide · 5 min readUGC hook ideas from the first 3 seconds of your own adsFind UGC hook ideas in your own ads: code the first three seconds of your strongest and weakest ads blind, then test new hooks on a body that works.
- Guide · 5 min readWrite UGC scripts that sound spokenWrite UGC scripts with a specific persona, connected thoughts and a visible product mechanism, then separate spoken input from caption text.
- Guide · 5 min readWhy AI ads feel uncanny: the details to inspectReview uncanny AI ad moments by gaze, skin, gestures, and voice cadence. Keep fictional speakers honest and replace shots that distract from the message.
- Guide · 5 min readLocalize video ads without translating the scriptLocalize video ads with locked facts, ordinary local phrasing, short pronunciation samples and captions timed to the approved language version.
- Guide · 5 min readCaptions for video ads: groups, timing and layoutBuild readable video ad captions with short meaning groups, restrained emphasis, accurate word timing, and a final phone-size layout review.
- Guide · 5 min read3D animation ads: build one coherent scenePlan a coherent 3D ad scene with a character bible, stable props and one action per shot, while keeping animation separate from product evidence.
- Guide · 5 min readCrochet-style AI ads: keep the world consistentKeep crochet-style AI ads coherent with locked stitch scale, material details, lighting changes, and frame-by-frame geometry checks before editing.
- Guide · 5 min readDrone explainer ads: labels that explain one ideaBuild drone explainers with stable architecture, controlled label anchors and phone-size reading checks, keeping illustrative views clearly scoped.
- Guide · 5 min readFaceless AI ads for brands: show the useful actionPlan faceless brand ads with objects, real demonstrations, diagrams, and animation. Give each shot an evidence role without an artificial spokesperson.
- Guide · 5 min readAI static ads: a clear image before decorationDesign AI-assisted static ads with clear visual hierarchy, editable typography, sourced claims and separately reviewed mobile crop versions.
- Guide · 5 min readMusic and sound effects for short video adsPlan short-ad sound effects around visible actions, protect speech, and audition the final music and effect mix on a phone speaker before delivery.
- Guide · 5 min readTalking-object ads: give the product a clear roleGive a talking object a consistent identity, purposeful speech and a matching voice, while separating character fiction from product promises.
- Guide · 5 min readHand an AI ad to an editor: what the folder needsPrepare an AI ad editor handoff with relinkable media, an edit recipe, captions, alternatives, and import checks. Separate recipes from interchange files.
- Guide · 5 min readUGC ad editing: the cut grammar we useEdit UGC ads around spoken thoughts, useful cutaways and clear cuts, then inspect caption timing and the rendered frames around every join.
- Guide · 5 min readZooms in video ads: punch-in, push and crashChoose video ad zooms by purpose, protect the speaker and product crop, and inspect keyframes for edge exposure, jitter, and caption obstruction.
- Guide · 6 min readA video approval workflow for small teamsSet up video approval with named decision owners, separate factual and style reviews, conflict resolution and a receipt for the exact final package.
- Guide · 5 min readFFmpeg for ad pipelines: traps from our rendersMaintain a Claude FFmpeg method with build checks, explicit layers, staged rendering and output inspection for version-sensitive production traps.
- Guide · 5 min readWrite a video ad brief an editor can act onWrite a video ad brief with a clear audience action, one message, real evidence, available assets, and acceptance checks an editor can use.
- Guide · 5 min readStoryboard an AI video ad before generating shotsStoryboard AI video ads with a beat-to-shot map, continuity anchors and required text layers, then check the plan against the approved voice.
- Guide · 5 min readWrite a product demo ad around a visible actionWrite a product demo ad script that follows a real visible action, states its result accurately, and leaves time for the viewer to understand each step.
- Guide · 5 min readEdit before-and-after footage with honest continuityEdit before-and-after footage with matched conditions, controlled labels and source notes, distinguishing real comparisons from staged illustration.
- Guide · 5 min readSpoken scripts vs captions: keep two text tracksKeep spoken scripts and captions as two mapped text tracks. Guide pronunciation in the voice input while preserving normal spelling and approved meaning.
- Comparison · 6 min readSRT vs WebVTT for video ad captionsCompare SRT and WebVTT syntax, cue maps, encoding and delivery modes, then test the exported captions in the receiver’s actual playback path.
- Guide · 5 min readFix caption drift from the first word to the lastDiagnose caption drift with opening, middle and ending measurements, distinguish offset from changing error and recheck the final audio version.
- Guide · 5 min readVertical video safe zones: review the final cropReview vertical video safe zones in the actual crop. Protect captions, faces, and product actions, and separate house framing rules from platform UI.
- Guide · 5 min readLoudness for social video: measure the finished mixMeasure social-video loudness and true peak separately, use a stated production target, and listen to speech, music, and the ending after the final encode.
- Guide · 5 min readMix music under voiceover without losing the wordsMix music under narration by checking arrangement density, steady levels and ducking, then audition ordinary devices and measure the final file.
- Guide · 5 min readColor space in video ads: keep overlays consistentKeep color consistent across video ad footage, graphics, and exports. Separate color metadata from pixel conversion, then inspect joins and real playback.
- Guide · 5 min readVideo ad export settings: check the delivered fileChoose export properties from a placement contract, inspect actual streams and review the delivered crop, captions and sound as one version.
- Guide · 6 min readDetect frozen frames in a video ad renderDetect unexpected stillness in video ad renders with timed review intervals. Separate freeze alerts, intentional holds, and black frames before repairing.
- Guide · 6 min readTrim and join ad clips with FFmpeg without driftTrim and join ad clips with FFmpeg using explicit ranges, reset timestamps, and matching streams. Verify cut frames, audio alignment, and final duration.
- Guide · 5 min readAdapt video ads to new aspect ratios shot by shotAdapt ad ratios shot by shot, reframe product actions, reflow captions and recheck graphic anchors before approving each complete delivery version.
Models
Video, image and voice models described by the job they do in an ad. Seedance, Kling, Veo and Hailuo for motion, Nano Banana for stills, ElevenLabs for the voice. Every model fact links to the maker’s own documentation, with the month we checked it. No winner and no prices.
Models on the Turntake site Coming
- Start here · 7 min readSeedance vs Kling vs Veo: which video model for which jobSeedance, Kling, Veo and Hailuo compared by the job in an ad, from b-roll to product shots, using each maker’s own docs. No winner and no prices.
- Model guide · 6 min readHow to use Seedance for ads: start frames and b-rollHow we use Seedance in ads: animate an approved start frame, one move and one action per shot, and what we do when it refuses a face. Dated tests.
- Model guide · 5 min readElevenLabs for ads: pace, pauses and pronunciationMake ElevenLabs voiceovers fit an ad: budget words to the length, write the script for the mouth, check pronunciation with speech-to-text, mix to −14 LUFS.
- Model guide · 6 min readNano Banana for ads: start frames and product stillsUse Nano Banana for the stills your ad starts from: product shots, the same person in new scenes and text in the market’s language. Our prompt rules.
- Guide · 5 min readSeedance face detected: the route from our testsHandle a Seedance face-detected refusal with a dated test record, a supported talking-head route, and separate b-roll. Keep the provider scope clear.
- Guide · 5 min readKeep the same person across AI ad shotsKeep an AI character recognizable across ad shots with an original portrait, a reference sheet, locked wardrobe, and angle checks before motion begins.
- Guide · 5 min readFix AI voiceover pronunciation before renderingFix pronunciation before animation with a model-specific dictionary check, a spoken script, a short audition, and a human review of the difficult words.
- Guide · 5 min readElevenLabs music for ads: plan the full music bedPrompt an ElevenLabs music bed around approved speech, leave an editable tail, and audition the ending before mixing the final short video ad.
- Guide · 5 min readSpeech to text for captions: time every wordTime captions against the approved voice with word timestamps, readable phrase groups, and full-track inspection. Keep spelling and timing checks separate.
- Model guide · 5 min readKling for ads: a shot test based on its docsPlan a Kling ad-shot test using its documented text and image inputs, a fixed reference, and a frame-by-frame acceptance sheet. No firsthand test claims.
- Model guide · 5 min readVeo for ads: plan shots and generated soundPlan a Veo 3.1 ad shot around its documented audio and frame inputs. Choose dialogue or separate narration, then review the picture and words separately.
- Model guide · 5 min readHailuo for ads: image-to-video shot planningUse MiniMax’s Hailuo documentation to plan an image-to-video ad shot, choose supported camera commands, and record a version-specific review before reuse.
- Guide · 5 min readLip-sync for dubbed ads: when the mouth must matchChoose between dubbing, revoicing, and regenerating an ad. Approve the audio first, inspect visible mouths, and keep every spoken claim supported.
- Guide · 5 min readAI video start and end frames: plan the shotPlan AI video start and end frames as compatible constraints. Check version support, inspect the motion between them, and keep a cut-based alternative.
- Guide · 5 min readWhy AI video text breaks and how to keep it legibleFix AI video text errors with clean source frames, separate overlays, and final-crop review. Preserve the text layer so the next edit stays correct.
- Guide · 5 min readAI video hands: inspect the action before motionInspect AI video hands before motion with a simple action, a clear contact point, and interval review. Keep product proof separate from illustration.
- Guide · 5 min readSet a voiceover word budget before the video cutSet a voiceover word budget from a real sample of the chosen voice. Reserve pauses and a readable close, then check each language against its actual take.
- Guide · 5 min readChoose an AI ad voice with a useful auditionChoose an AI ad voice with the same sample script, pronunciation checks, fluent review, and a picture pairing. Keep preference separate from evidence.
- Guide · 5 min readFix voiceover pacing without breaking the editFix voiceover pacing by choosing between a text rewrite, careful pause edits, and beat placement. Rebuild word timing and inspect the final voice-led cut.
Claude
How we run the factory with Claude: skills that hold the house rules, connections to the models over MCP, and a person who presses go. Claude is a product of Anthropic, and Turntake isn’t affiliated with Anthropic.
- Start here · 7 min readHow to use Claude for ads: skills, MCP and guardrailsHow we run ad production with Claude: skills for the method, MCP for the tools, hard rules for keys, budgets and Meta, and what stays human.
- Guide · 5 min readCan Claude edit video ads? Render from an edit fileEdit existing video ads with Claude from a versioned recipe: define limited operations, preserve required layers and review the rendered result.
- Guide · 5 min readConnect ElevenLabs to Claude for ad voiceoversPlan an ElevenLabs connection for Claude, inspect its tools, protect credentials and approve one pronunciation sample before a voiceover batch.
- Guide · 5 min readCan Claude run Meta ads? Our read-only boundaryUse Claude for read-only ad analysis, preserve recommendation evidence and keep uploads, paused staging and activation with a responsible person.
- Comparison · 6 min readClaude skills vs MCP for ad productionCompare Claude skills, MCP and subagents for ad work: choose written method, external access or bounded workers, with clear review boundaries.
- Guide · 5 min readWrite a Seedance prompt skill for ClaudeBuild a Seedance prompt skill around an approved reference, locked visual details, one action and a concrete input-output acceptance example.
- Guide · 5 min readPlan Nano Banana stills with ClaudePlan Nano Banana stills with Claude using image briefs, reference roles and file-specific approval before passing the selected frame to motion.
- Guide · 5 min readRun AI agents on an ad batch: briefs and boundariesDivide an ad batch into isolated agent assignments with fixed inputs, hard limits, written acceptance criteria and a reviewable merge manifest.
- Guide · 5 min readA checker that does not fix: review AI ad outputsDesign an independent AI review pass that counts files, locates defects, preserves submitted work and rechecks the actual revised output.
- Guide · 5 min readAI project memory: decisions in files, not chatOrganize AI project memory with an index, durable rules, current tasks and dated decisions, then test whether a new session can follow the handoff.
- Guide · 5 min readA video edit JSON recipe that survives revisionDefine a video edit JSON contract with exact assets, time units, ordered layers and required graphics, then bind the recipe to its final hash.
- Guide · 5 min readVersion control for video ads: know the approved cutTrack video ads with immutable outputs, recipe differences and scoped approval manifests so notes and decisions identify the actual reviewed cut.
- Guide · 5 min readBuild a video render queue with visible job statesDesign a render queue with visible states, atomic job claims, bounded concurrency and validated temporary outputs before marking work complete.
- Guide · 5 min readValidate AI video ad outputs before human reviewValidate ad recipes, media, captions and required graphics with located failure reports, then keep mechanical success separate from editorial approval.
- Guide · 5 min readRetry AI generation jobs without submitting twiceRecover generation jobs from recorded request IDs, separate uncertain submissions from failures and verify downloaded assets before another attempt.
Your own footage
Most brands already have the footage for their next ads. These guides start from the ads that worked, find the clean footage behind them and sort it into a library of clips, with AI b-roll only where it is honest.
Bring your own content on the Turntake site Early access
- Start here · 7 min readAI ad workflow for small business: start from your footageA workflow for small brands with ad history: find your best ads, trace the clean footage, sort clips by job, then film or generate only what’s missing.
- Guide · 5 min readHow to organize video files for editing adsOrganize video footage by the job each clip does in an ad: map every file first, name clips by what they show and keep talking heads apart from b-roll.
- Guide · 5 min readHow to repurpose video ads: find the clean footageRepurpose video ads by starting from the ones that worked and tracing the original footage behind them, before captions were burned in. Match by picture.
- Guide · 6 min readAI b-roll for product ads: what to generate, what to filmGenerate the mood, film the proof. Where AI b-roll works in a product ad, where it doesn’t, and the four checks every clip passes before it goes in an ad.
- Guide · 5 min readTag a video clip library with AI: rules firstTag a video library with controlled fields, separate visible speech from b-roll, retain uncertainty and audit labels against the actual clips.
- Guide · 5 min readDo you have enough footage for new video ads?Assess whether your footage can support new ads with a coverage matrix for openings, explanation, proof, and close. Turn missing roles into a filming plan.
- Guide · 5 min readFind hooks and selects in editor project filesUse editor project selections to locate opening footage, retain source ranges, and understand creative order. Keep selection separate from permission.
- Guide · 5 min readResearch your ad library through read-only accessResearch ads through traceable source maps, opening comparisons and production hypotheses, keeping public observations and owned reports separate.
- Guide · 5 min readLog footage in ad edits: transform or set asidePrepare log footage for ad edits by identifying its camera profile, applying the right technical transform, and reviewing color before creative styling.
- Guide · 5 min readBuild a product video shot list from proof needsBuild a product shot list from proof needs, capture purposeful wide, medium and close views, and inspect the real actions before editing.
- Guide · 5 min readFilm phone b-roll that is useful in an ad editFilm phone b-roll with stable framing, controlled focus and exposure, complete actions, and extra handles. Review the actual ad crop before leaving.
- Guide · 5 min readVariable frame rate footage in a video ad editPrepare variable frame-rate footage with timestamp inspection, preserved originals and a deliberate intermediate, then compare duration and sync.
- Guide · 5 min readDeduplicate a video library while keeping provenanceDeduplicate footage by separating exact bytes, similar takes and technical derivatives, preserving provenance and usage before reviewed archiving.
- Guide · 5 min readName video assets with stable IDs and readable labelsName video assets with stable IDs and readable labels. Track versions, source relationships, and renames so footage stays findable across projects.
- Guide · 5 min readVideo review proxies: small files, traceable notesCreate review proxies with explicit source mapping, faithful crops and tested note timing, then verify important decisions on the full-quality final.
- Guide · 5 min readArchive video ad assets without losing the next editArchive video ad assets as a restorable package: approved export, editable recipe, source media, captions, audio, dependency records and a reopening check.
Testing and scaling
What happens after an ad goes live: creative fatigue, one change per test, a new hook on a body that already works, and a b-roll library that lets you vary ads without reshooting.
- Start here · 7 min readCreative fatigue on Meta: new ads without reshootingWhat creative fatigue is on Meta, and how to make different ads without a new shoot: new hooks, new formats and a clip library with a reuse limit.
- Guide · 5 min readHow to iterate winning ads: new hooks, same bodyLock the body of a winning ad and change only the hook, city or language. Edit the body once, time each hook to its own audio and check every version.
- Guide · 5 min readCreative testing on Meta: one change per testOne change per test, a written guess before the ad is made, pretests as a filter, and a human who presses go on a paused ad. Creative testing, kept simple.
- Guide · 5 min readB-roll library for ads: reuse clips without repeatsBuild a b-roll library you can reuse: cut clips at real boundaries, save them silent with a clear index, pick the least-used match and log every use.
- Comparison · 6 min readHook rate vs hold rate: define what you measureCompare hook, thumbstop and hold measures by defining each event, denominator and reporting window before drawing a bounded creative conclusion.
- Guide · 5 min readPretest ads with synthetic personas: compare pairsCompare ad pairs with synthetic personas, balanced order and raw reasons, then turn useful clarity findings into hypotheses for observed tests.
- Guide · 5 min readHow many ad creatives? Plan useful variety firstPlan useful ad creative variety around distinct ideas, available test slots, and team capacity. Keep one hypothesis clear instead of chasing a fixed count.
- Guide · 5 min readMeta Andromeda: what creative variety changesRead Meta’s Andromeda research as retrieval-system context, then plan distinct creative ideas without turning engineering facts into a reach promise.
- Guide · 5 min readCity and language versions of one approved adBuild city and language ad variants from verified local facts and an approved master, preserving the fixed body and reviewing each changed hook.
- Guide · 5 min readKeep an ad test hypothesis log you can act onKeep an ad test log with a written hypothesis, fixed asset versions, metric definitions, delivery context and a clear decision for the next creative round.
- Guide · 5 min readCreative test confounders: when a result is unclearCheck ad creative test confounders before choosing a winner: audience, timing, placement, offer, tracking, asset versions and unequal delivery.
- Guide · 5 min readPlan clip reuse across ad concepts without repetitionPlan clip reuse by visual meaning and concept, track intentional recurrence and identify precise new-shot gaps instead of counting filenames.
How we write these
- No invented numbers, prices or quotes. If we didn’t count it in our own production or read it in a source, it isn’t here.
- Model facts come from the maker’s own docs, linked in the text, with the month we checked them. Models change, so read the linked page again before you plan a batch.
- We don’t rank models. We describe the job each one did for us and the spec we tested it against. A model that failed our spec on one day isn’t a bad model.
- Our observations carry a date, like “in our tests, October 2026”, because they can go out of date. The date sits in the sentence itself or at the top or foot of its section or table.
- What isn’t built yet says Coming. What we set up with you says Early access.
- You press go. The guides describe how we make ads, not what you should publish. Nothing in our own work goes live without a person deciding it.