product-ad-production — скилл для Claude Code
Скилл product-ad-production для Claude Code проектирует, сценарирует и проверяет видеорекламу продуктов на базе доказательств.
MIT 177 установок ★ 252 обновлён 2026-07-14
npx skills add calesthio/generative-media-skills --skill product-ad-production -a claude-codeЧто делает product-ad-production
Скилл диагностирует коммерческую задачу и выстраивает систему утверждений для рекламных роликов перед созданием визуального ряда. Он проверяет бриф по четырём категориям готовности и помогает сформировать требования к доказательствам до написания сценария.
Инструмент помогает учитывать правовые требования к рекламе и формировать реестр утверждений со стабильными идентификаторами. Он работает с физическими товарами, ПО, приложениями, услугами и цифровыми продуктами.
Какие задачи решает
- составить бриф для рекламного видеоролика
- проверить доказательную базу для маркетинговых утверждений
- построить реестр рекламных тезисов с привязкой к фактам
- диагностировать коммерческую проблему перед запуском рекламы
Примеры запросов агенту
- Оцени бриф продукта по критериям готовности
- Создай реестр утверждений для рекламного ролика
- Разработай сценарий видеорекламы для приложения
Как установить product-ad-production
1Через Skills CLI (Vercel)
npx skills add calesthio/generative-media-skills --skill product-ad-production -a claude-code2Вручную: положить файлы
~/.claude/skills/product-ad-production/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/product-ad-production/
1Через Skills CLI (Vercel)
npx skills add calesthio/generative-media-skills --skill product-ad-production -a codex2Вручную: положить файлы
.agents/skills/product-ad-production/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add calesthio/generative-media-skills --skill product-ad-production -a cursor2Вручную: положить файлы
.cursor/skills/product-ad-production/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install calesthio/generative-media-skills skills/production/content-formats/product-ad-production2Вручную: положить файлы
.github/skills/product-ad-production/SKILL.md1Через Skills CLI (Vercel)
npx skills add calesthio/generative-media-skills --skill product-ad-production -a gemini-cli2Командой платформы
gemini skills install https://github.com/calesthio/generative-media-skills --path skills/production/content-formats/product-ad-productionВручную: положить файлы
.devin/skills/product-ad-production/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add calesthio/generative-media-skills --skill product-ad-production -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о product-ad-production
Как установить product-ad-production в Claude Code?
Выполните команду npx skills add calesthio/generative-media-skills --skill product-ad-production -a claude-code, а затем используйте путь ~/.claude/skills/product-ad-production/SKILL.md.
Что делает скилл product-ad-production?
Он помогает диагностировать коммерческие задачи, выстраивать систему доказательств и создавать сценарии видеорекламы на основе проверенных фактов.
На чём основана проверка рекламных тезисов?
Проверка опирается на документированные факты регуляторов, эмпирические результаты исследований и рабочие правила производства.
Исходный текст
SKILL.md · MIT
--- name: product-ad-production description: Diagnose, design, script, produce, adapt, and quality-control product video advertisements from evidence-backed briefs. Use for physical-product, software, app, service, ecommerce, direct-response, brand-response, demonstration, testimonial, launch, and variant ad work where product truth, legibility, claims, platform delivery, and measurable iteration matter. --- # Product Ad Production Produce an advertisement that makes the right person recognize a relevant situation, understand what the product changes, believe the proof, and know what to do next. Visual polish is not a substitute for product comprehension or claim support. This skill is provider-neutral. Select image, video, capture, voice, music, editing, and composition tools only after the advertising problem is defined. Do not let a generation model's favorite visual language determine the concept. ## Evidence language used here - **Documented fact** means a requirement or principle stated by a regulator, standards body, or platform source. - **Research finding** means an empirical result with a stated source and scope. It is evidence, not a guarantee for a new campaign. - **Practitioner heuristic** means a production rule worth testing, not a universal truth. Platform behavior and ad products change. Platform facts in this skill were verified **2026-07-09**. Re-check the exact placement specification, policy, interface preview, and destination requirements immediately before export and trafficking. ## Start with a decision-grade brief Do not write a script from a slogan and a product URL. Diagnose the commercial problem first. ### Minimum input record Capture or infer each field, marking assumptions explicitly: | Field | Required decision | |---|---| | Business outcome | Awareness, consideration, lead, trial, purchase, upsell, retention, or launch learning | | Primary success measure | One decision metric; supporting and guardrail metrics are separate | | Audience | Purchase situation and constraints, not demographics alone | | Job to be done | Progress sought, present struggle, emotional/social stakes, and competing solutions | | Awareness state | Unaware, problem-aware, solution-aware, product-aware, or returning customer | | Product truth | What the product is, how it works, prerequisites, limitations, price/offer, and current availability | | Proposition | One useful change the ad will make credible | | Evidence | Approved substantiation, demonstrations, customer data, certifications, testimonials, and provenance | | Objection | The most consequential reason a qualified viewer may not act | | Placement | Platform, ad product, buying objective, duration, aspect ratio, UI overlays, and sound context | | Destination | Exact landing page or store state; product, offer, naming, and price must agree with the ad | | Brand system | Correct product/packaging, marks, type, colors, sonic assets, and prohibited treatments | | Rights and policy | Talent, location, music, stock, creator, user content, AI/replica consent, category restrictions | | Production constraints | Deadline, budget, available real product, UI build, languages, render and review path | Return one of three diagnoses: - **Ready:** proposition, evidence, destination, and rights can support production. - **Conditional:** production can proceed with named assumptions or conservative wording pending approval. - **Blocked:** a material claim, offer, product state, destination, or permission is missing. Do not invent it. ### Audience through the job, not the persona costume Jobs to Be Done theory treats products as choices people make to achieve functional, emotional, and social progress, not merely as bundles of attributes [S1]. Translate the job into an ad-specific record: ```text When [situation/trigger], this person is trying to [progress], but [friction/anxiety/current alternative]. They will believe a better option when they see [proof], and the smallest credible next action is [CTA]. ``` Separate buyer, user, approver, and beneficiary when they differ. For B2B, map the operational user, economic buyer, security/legal reviewer, and implementation owner. An ad normally speaks to one primary role and acknowledges one decisive objection; it should not average all roles into vague corporate language. ## Build a claim system before a visual system **Documented fact—United States:** advertising must be truthful and non-deceptive, advertisers must possess evidence for claims, and ads cannot be unfair [S2][S3]. The FTC evaluates the ad's overall or “net” impression, including words, images, sounds, omissions, and format; a small disclosure does not neutralize a misleading main message [S4]. Other jurisdictions and regulated categories may impose additional rules. Obtain qualified legal review when stakes warrant it. ### Claim ledger Create the ledger before scripting. Every spoken, written, visual, sonic, comparative, and implied claim gets an ID. | Field | Record | |---|---| | Claim ID | Stable identifier used in script, shot list, and approvals | | Exact expression | Wording or visual implication the viewer receives | | Type | Existence/feature, performance, quantified, comparative, superiority, price/offer, testimonial, environmental, health/safety, or illustrative | | Audience interpretation | Reasonable takeaway, including likely implied meaning | | Evidence | Document, dataset, test report, product record, or approved customer source | | Scope | Model/SKU, population, conditions, geography, dates, and exclusions | | Qualification | Material information needed beside the claim | | Demonstration link | Shot IDs that prove or illustrate it | | Owner and approval | Product/legal/brand approver and status | | Expiry trigger | Product revision, offer end, evidence age, or policy change | ### Evidence discipline 1. Prefer direct, current evidence for the exact advertised product and conditions. 2. Match the claim's strength to the evidence. A measured customer outcome is not automatically a causal product effect; a laboratory result is not automatically typical consumer performance. 3. Preserve denominators, sample, comparison baseline, time period, and material conditions for quantified claims. 4. Use “up to,” “from,” or “as much as” only when accurate and not misleading about expected results. 5. Treat the depiction itself as a claim. An impossible cleanup, instant software response, enlarged product, synthetic testimonialist, or hidden edit can create a false impression without a false sentence. 6. If a scene is conceptual rather than evidentiary, make its illustrative role unambiguous. Do not style an animation like a measurement readout unless it reflects measured data. 7. Keep required qualifications proximate, legible, understandable, and present long enough to read. Rewrite the main claim when the qualification would need to contradict it. 8. Never fabricate awards, press marks, certifications, reviews, ratings, customers, interfaces, prices, scarcity, before/after states, or competitor behavior. ### Testimonials, creators, and reviews **Documented fact—United States:** endorsements must reflect the endorser's honest experience; material connections should be clearly disclosed, and atypical results may require disclosure of generally expected results [S3][S5]. The FTC advises placing a social disclosure with the endorsement itself, making it hard to miss, and including it in the video rather than relying only on a description [S5]. The U.S. Consumer Reviews and Testimonials Rule prohibits specified fake or false reviews and testimonials, including certain AI-generated fakes [S6]. Maintain a testimonial packet: original statement, permission, identity, actual-use confirmation, compensation/material connection, editing approval, validity date, and support for any product claim repeated by the endorser. An actor may portray a dramatized user only when the ad cannot reasonably be read as that actor's genuine testimonial. Label dramatizations where needed. ## Choose a product-centered concept Compress the strategy into: ```text For [audience in situation], [product] changes [friction] into [outcome] because [credible mechanism/proof]. After viewing, do [one action]. ``` Generate multiple concepts that differ in mechanism, not merely art direction. Useful concept families include: - **Observable transformation:** a credible before/use/after with controlled continuity. - **Mechanism reveal:** show the feature causing the benefit. - **Stress or boundary test:** demonstrate a supported limit under disclosed conditions. - **Routine compression:** contrast the old sequence with the product-enabled sequence without manipulating elapsed time. - **Objection reversal:** surface the barrier, then prove the answer. - **Evidence story:** build around a verified user or business result with scope visible. - **Product-as-interface:** let the real workflow, handling, sound, or result carry the message. - **Identity or ritual:** associate product use with an emotional/social job while keeping the product causally legible. Score each concept from 0–3 on audience relevance, claim support, product centrality, demonstration clarity, brand distinctiveness, placement fit, variant potential, feasibility, and rights/policy risk. Reject any concept whose appeal depends on a behavior the product cannot perform. **Research finding—platform-specific:** Google's data-backed ABCD guidance recommends earning attention, branding early and throughout, creating a focused human connection, and giving clear direction; for consideration it specifically recommends making the product prominent and showing how it works [S7]. Google's reference guide also says its evidence is correlational and not a one-size-fits-all guarantee [S8]. Use this as informed prior evidence, then test it for the brand and placement. ### Product legibility contract The viewer should be able to answer, at the appropriate moment: 1. What is being advertised? 2. What does it do or change? 3. How do I know? 4. Is this the exact product/plan/offer I will reach? 5. What should I do next? Design explicit recognition events: - an early product or brand cue appropriate to the objective; - at least one unobstructed, correctly shaped product/package/UI view; - a readable name or a distinctive asset that the intended audience can actually link to the brand; - product interaction at a scale that makes the mechanism understandable; - a late memory/choice cue and destination-consistent CTA. Do not assume that a tiny logo, generic color, or beauty silhouette identifies an unfamiliar product. The Ehrenberg-Bass Institute describes a strong distinctive asset as both famous and unique to the brand; an unmeasured visual element should not be treated as a brand-name substitute [S9]. **Practitioner heuristic:** perform a “blurred-thumbnail test” for silhouette/color recognition, a one-second freeze-frame test for product/category recognition, and a five-second recall check with people who did not see the brief. These are diagnostic screens, not campaign-effectiveness measures. ## Design demonstrations as evidence A demonstration is a small experiment presented to viewers. Define it before shooting or generating: | Element | Question | |---|---| | Claim | What exact viewer conclusion may this demo support? | | Unit | Which real SKU, version, plan, build, and accessories are used? | | Conditions | What environment, starting state, preparation, network, lighting, load, or operator skill matters? | | Comparator/control | Is the baseline fair, current, and equivalently prepared? | | Procedure | What actions happen, in what order, and how are they timed or measured? | | Capture | Which uninterrupted wide proof shot and which detail inserts are needed? | | Repetitions | Is one event enough, or must variability/typicality be shown? | | Qualification | Which material conditions must appear with the result? | | Failure rule | What result stops the advertised claim or requires revised language? | For physical products, capture a continuity master that shows setup, product, hands, comparator, timer or measuring device, and outcome without a claim-altering cut. Detail shots can improve comprehension but cannot replace the proof master. For software and apps: - use the approved production build or a clearly marked prototype; - preserve real labels, steps, latency, permissions, data states, and prerequisites; - use synthetic data that cannot expose private or confidential information; - do not remove required steps merely to make the workflow seem instant; - distinguish a UI concept animation from an actual screen recording; - verify the landing experience matches the plan, feature, and availability shown. For comparisons, match conditions and show the basis of comparison. Avoid degrading a competitor, choosing an obsolete baseline without disclosure, changing camera treatment to favor one outcome, or saying “better” without defining the dimension. **Practitioner heuristic:** if the proof cannot survive a locked-off wide shot, it is not yet a demonstration; it is an illustration. Illustration can still be useful, but label and script it accordingly. ## Turn the proposition into time There is no mandatory hook-problem-solution template. Build beats around the viewer's decision and the placement's interruption pattern. Possible beat functions: - **Pattern entry:** a relevant action, consequence, contradiction, question, or result—not unrelated spectacle. - **Situation recognition:** make the viewer recognize their job and friction. - **Product identification:** name and show the product without delaying until attention is already spent. - **Mechanism or use:** reveal how the feature changes the situation. - **Proof:** demonstrate, quantify with scope, or present an authentic experience. - **Objection handling:** answer the barrier that would otherwise block action. - **Consequence:** show the useful or emotional result without overclaiming causality. - **Choice and action:** reinforce product, offer if applicable, and one next step. ### Beat sheet For every beat, record: ```text time range | beat purpose | viewer question answered | visual action product/brand cue | dialogue/VO | on-screen text | claim IDs audio event | transition logic | source asset | disclosure/qualification placement risks | cutdown dependency ``` ### Script rules - Write for the spoken duration and leave comprehension room; do not fill every frame with words. - Let image, speech, text, and sound contribute different information. Do not make four channels compete. - Keep the primary proposition expressible in one sentence. - Use product language the audience uses, but retain exact terms where plan/version differences matter. - Bind every number and superlative to a claim ID. - Make qualifications syntactically understandable, not a legal-text dump. - Ensure the CTA follows sufficient context. “Shop now” before the viewer knows what or why is not direction. - Avoid fake urgency, clickbait that the body cannot repay, and first-person claims spoken by synthetic talent as if personally experienced. ### Shot architecture Specify shots by advertising function, not by decorative camera vocabulary: | Shot role | Acceptance criterion | |---|---| | Product identity | Exact product is recognizable; logo/label is not malformed or obscured | | Situation | Friction is clear without stereotyping or humiliating the user | | Operation | Contact points, sequence, orientation, and causality are readable | | Proof master | Claim-relevant action remains continuous and conditions are visible | | Detail | Reveals a mechanism or result the master cannot show | | Human consequence | Performance and result match the intended audience/job | | Offer/CTA | Product, offer, destination, terms, and next action agree | AI-generated shots require additional continuity controls: approved pack/reference images, geometry and label verification, state tracking across shots, hand/product interaction review, and a ban on hallucinated ports, controls, accessories, ingredients, results, or interface elements. When accuracy is essential, use real product capture, composited pack shots, deterministic 3D, or real UI rather than asking a generative model to reproduce critical details. ## Variant design is hypothesis design Do not call alternate color grades “three concepts.” Build a variant matrix where each version has one named learning goal. | Variable | Example hypothesis | |---|---| | Situation/hook | Recognized morning friction will improve qualified hold versus abstract intrigue | | Audience/job | Team lead outcome language will outperform agent-level feature language for lead quality | | Proposition | Time saved will outperform compactness for commuters | | Proof | Continuous demonstration will outperform testimonial-only proof | | Objection | “No setup required” will increase trials among switchers | | Offer | Trial framing will outperform discount framing on qualified conversion | | CTA | “Build your first board” will outperform generic “Learn more” | | Length/placement | A native 12-second proof cut will outperform a cropped 30-second master on vertical feed | Create a version manifest with parent version, changed variable, preserved variables, claim-ledger version, assets, platform, language, duration, export checksum, approval state, and live dates. **Documented platform method:** TikTok's split-testing guidance describes isolated audience groups and holding other variables constant; Google Video Experiments can compare different video ads with the same audience [S10][S11]. Platform systems and eligibility differ, so plan enough traffic/time and use the platform's experiment design rather than judging early noise. **Practitioner heuristic:** begin with materially different propositions or proof approaches, then optimize execution within the winning strategic territory. Tiny changes too early produce precise answers to unimportant questions. ## Audio must carry meaning, not merely energy Plan four layers: 1. **Speech:** intelligible performance, natural cadence, correct product pronunciation, and no unsupported personal experience. 2. **Product sound:** real or truthfully designed interaction sounds that clarify mechanism and physicality. 3. **Music:** licensed mood, momentum, and brand character that never masks claims or disclosures. 4. **Sonic identity:** an owned, approved cue used consistently enough to aid recognition. Maintain a cue sheet with track/source, composer/performer, license, territories, media, term, paid-social rights, edits, stems, and expiry. “Royalty-free” does not mean unrestricted. Confirm whether a platform music library permits paid advertising and off-platform versions. Caption all meaningful speech and relevant non-speech information when accessibility or destination context calls for it. **Documented standard:** WCAG 2.2 requires captions for prerecorded synchronized media in covered web contexts; captions should not obscure relevant visual information [S12]. For critical text, target WCAG's 4.5:1 contrast for normal text and 3:1 for large text where applicable [S13], while recognizing that platform video itself may be governed by different formal requirements. Mix against the destination's current specification. Check dialogue intelligibility, true peak/clipping, phase, mono compatibility, mobile-speaker translation, and head/tail silence. For U.S. digital television delivery, ATSC A/85 is the relevant loudness recommended practice; do not apply a broadcast target blindly to social exports [S14]. **Practitioner heuristic:** the visual track should communicate product, change, and CTA with sound unavailable; the audio track should still identify product and proposition when the viewer glances away. This is a robustness test, not an instruction to weaken either channel. ## CTA and destination continuity A strong CTA names one feasible action and the immediate value of taking it: - “See colors” is better than “Learn more” when the landing page is a color selector. - “Start a 14-day trial” is appropriate only if the trial, length, eligibility, and billing terms are accurate. - “Book a workflow review” should land on that booking flow, not a generic homepage. Verify product/SKU, brand name, price, discount math, dates, inventory, shipping territory, eligibility, subscription renewal, button text, app-store state, language, tracking parameters, and page speed/function. TikTok's current policy explicitly requires consistency among ad, caption, CTA, product, and landing page [S15]; treat this as good production discipline everywhere. ## Adapt for placements; do not merely crop Build a clean, high-resolution master with separated dialogue, music, effects, captions, graphics, textless plates, and product/pack layers. Then author placement versions. For each placement: 1. Retrieve the current official spec for the exact ad product and buying objective. 2. Reframe action and eyelines for the aspect ratio; do not auto-center blindly. 3. Re-time the opening, brand cue, proof, and CTA for the skip/swipe behavior. 4. Move product, captions, disclosure, price, and CTA away from dynamic UI overlays. 5. Rebuild type to remain readable on the intended device. 6. Use permitted audio only; licensed organic music may not be eligible for ads. 7. Preview inside the platform tool on representative devices. 8. Export and watch the uploaded/transcoded result, not only the local master. ### Current placement anchors, verified 2026-07-09 These are planning anchors, not a substitute for the live spec: - **YouTube Shorts:** Google recommends vertical 9:16, sound-on/social-first assets and generally recommends less than 60 seconds for Shorts behavior; the exact eligible duration depends on campaign/ad format [S16]. - **Meta Reels:** Meta recommends 9:16 video with audio and key messages in the safe zone, and provides a safe-zone checker/preview path [S17]. Instagram notes that Reels ads may be up to 15 minutes but recommends full-screen vertical and original or royalty-free audio rather than licensed music [S18]. Do not confuse maximum eligibility with an effective duration. - **TikTok in-feed:** current auction guidance recommends vertical 9:16 at at least 540×960, allows other ratios, and makes safe zone depend on dimensions, caption length, and added formats [S19]. A separate current policy page states standard ratios, minimum/maximum duration, audio, legibility, and dynamic-content requirements for covered markets [S15]. Because TikTok documents differ by ad product, use the exact placement page and preview tool [S20]. - **YouTube in-stream/bumper:** skippability and duration differ by format; Google's current help identifies bumper ads as six seconds or less and skippable in-stream as skippable after five seconds [S21]. Author the message for the purchased format rather than sending the same cut everywhere. Create a delivery manifest containing placement, ad product, objective, duration, canvas, pixel dimensions, frame rate, codec/profile, bitrate or quality mode, audio codec/sample rate/channels, loudness target if specified, captions mode/file, safe-zone template version, filename, checksum, source master, claims version, and approval. ## Measurement and iteration Tie evaluation to the business decision: - **Awareness:** incremental reach, valid/viewable delivery, attention/view measures with known definitions, brand lift, correct brand/product linkage. - **Consideration:** qualified view behavior, product-page engagement, search lift, feature exploration, lead quality. - **Action:** incremental conversion where possible, conversion rate, cost per qualified action, contribution margin/ROAS with attribution caveats, cancellation/return guardrails. - **Experience guardrails:** negative feedback, hide/report rate, misleading-click signals, refund/return reasons, complaint themes, policy rejection. The MRC viewability convention counts a viewable video impression when at least 50% of pixels are in view for at least two continuous seconds [S22]. That is a measurement threshold, not proof that the ad was noticed, understood, remembered, or persuasive. Use a learning log: ```text hypothesis | treatment/control | single changed variable | audience/placement primary metric | guardrails | planned sample/time | result/uncertainty claim or product changes during test | decision | next test ``` Do not optimize a brand objective solely to click-through rate, or a sales objective solely to completion rate. Do not declare a winner from a few conversions, overlapping audiences, unequal offers, mid-test edits, or a platform label you do not understand. Where feasible, use randomized experiments or lift studies; Google distinguishes A/B experiments from exposed-versus-control lift studies for measuring incremental outcomes [S23]. ## Quality gates ### Before asset generation or capture - Brief has one business outcome, one primary audience/job, one proposition, one objection, and one CTA. - Claim ledger is approved or conservatively scoped; blocked claims are absent. - Product/SKU/build, offer, and destination are frozen for the version. - Concept passes product centrality, proof, feasibility, and policy review. - Demonstration protocol, continuity master, UI state, and failure rule are specified. - Rights plan covers every person, place, mark, asset, recording, voice, and AI replica. ### Rough-cut review - A cold viewer can identify product/category, change, evidence, and next action. - Every claim maps to evidence and appears with required scope. - Product geometry, packaging, controls, UI, prices, and outcomes are accurate. - Proof is not created by a hidden cut, time compression, selective framing, or impossible animation. - Brand cues are attributable; creator/advertising disclosures are hard to miss. - Captions, disclosures, and qualifiers are readable and do not collide with UI. - Speech is intelligible; music/effects support rather than obscure. - No scene depends on an unlicensed asset or unauthorized likeness/voice. ### Final technical and traffic review - Watch once at normal speed on target-size mobile and once frame-by-frame. - Review mute, headphones, phone speaker, and mono; check clipping and sync. - Inspect first frame, skip point, thumbnail if used, CTA frame, and loop boundary. - Check safe zones in official preview; simulate longest localized text. - Confirm codec, dimensions, frame rate, duration, file size, audio, captions, and filename against the live spec. - Upload a proof and inspect platform transcode, text crop, caption timing, color, and destination. - Archive master, textless, stems, captions, claim ledger, rights evidence, approvals, platform previews, and delivery manifest. ## Common failures and repairs | Failure | Why it fails | Repair | |---|---|---| | Beautiful category film; product appears at end | Attention is not linked to the advertised choice | Put product/mechanism into the premise and create early recognition | | Feature inventory | No prioritized job or consequence | Choose one situation, one change, one proof | | Generic “game-changing” language | Unsupported and non-diagnostic | State the exact outcome and show evidence | | Demonstration built from inserts only | Cuts can conceal setup or causality | Add an uninterrupted proof master and disclose material conditions | | Fake interface speed | Implies performance not delivered | Capture real latency or label a concept; edit around non-material pauses only | | Testimonial by synthetic person | Can be read as a real experience | Use a documented real endorser or unmistakable dramatization | | Tiny disclaimer contradicts headline | Net impression remains misleading | Narrow or remove the headline claim | | One master auto-cropped everywhere | Product/text/UI are occluded; pacing mismatches placement | Re-author framing, type, beats, and CTA per placement | | Thirty variants with no hypothesis | No interpretable learning | Create a version matrix and isolate consequential variables | | Optimizing to cheap views | Delivery proxy displaces business goal | Predeclare primary and guardrail metrics | | AI-mutated pack/UI between shots | Product becomes fictional | Lock approved references; composite or use real/3D/UI capture | | Music chosen after edit | Rights, rhythm, and intelligibility break late | Decide source/license and audio role before picture lock | | CTA and landing page disagree | Viewer trust and platform compliance suffer | QA the live destination with every trafficking version | ## Rights, safety, and responsible production - Secure written rights for product marks, packaging, stock, locations, artwork, fonts, music, performances, creator content, testimonials, customer logos/data, and usage territories/media/term. - Obtain explicit, scoped consent for a person's face, voice, performance, and any digital replica. Do not infer permission to clone from permission to record. - Never create a fake consumer, expert, celebrity, employee, or creator endorsement. Do not imply real product use by someone who did not use it. - Minimize personal/confidential data in screen capture; use approved synthetic accounts and revoke exposed credentials. - Preserve truthful body appearance and product effects. Avoid exploitative fear, discriminatory targeting/stereotypes, unsafe demonstrations, and instructions likely to cause harm. - Treat products involving health, finance, credit, housing, employment, alcohol, gambling, children, politics, environmental benefit, safety, or regulated performance as specialist-review categories. - Check each platform's current advertising and synthetic-media policies, not only technical specs. - Keep an AI provenance reco...
© calesthio · MIT · calesthio/generative-media-skills