Communitygithub.com

jiawood2006/hermes-skills

AI 电商视频广告生产线:一条命令出带货短视频(产品锁 / 多引擎选型 / 官方真像素产品镜 / 自动装配)。One-command product ad video pipeline: product lock, multi-engine selection, official-pixel product inserts, auto overlay assembly.

hermes-skills 是什么?

hermes-skills is a Claude Code agent skill that aI 电商视频广告生产线:一条命令出带货短视频(产品锁 / 多引擎选型 / 官方真像素产品镜 / 自动装配)。One-command product ad video pipeline: product lock, multi-engine selection, official-pixel product inserts, auto overlay assembly.

兼容平台~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/jiawood2006/hermes-skills/tree/HEAD/skills/ai-video-ad-production

在你喜欢的 AI 中提问

打开一个已预加载此 Agent Skill 的新对话。

文档


name: ai-video-ad-production description: AI e-commerce video ad production line: one-command product ad video pipeline (product lock, multi-engine selection, official-pixel product inserts, auto overlay assembly). category: ecommerce

AI-generated short product videos (live-action story / voice-over)

When to Use

When the user asks for "short video / promotional video / voice-over video / sample standard", or complains that previous results "lack realism, look like image collages". First, make a path decision (see below) — do not treat image stitching as a deliverable.

🚀 How to Run (As of 2026-09-18: One Command to Generate Video)

Main command = scripts/make_ad.py — consolidates the previous 6 manual scripts scattered in /tmp into a single command. All subsequent component scripts are called by it as needed.

S=~/.hermes/skills/ecommerce/ai-video-ad-production

# ① First, preview the prompt (free, no submission; confirm product lock + storyboard before spending money)
python3 $S/scripts/make_ad.py $S/templates/ad_project.hsq1.json --steps plan

# ② Fully automated one-stop: submit → poll → download → frame extraction verification → assembly 30s/15s → deliver to desktop
python3 $S/scripts/make_ad.py $S/templates/ad_project.hsq1.json --steps all

# ③ Segments already downloaded, only re-run post-production (change subtitles/voiceover/re-cut to 15s — no cost, most common use)
python3 $S/scripts/make_ad.py <project.json> --steps assemble,deliver

# ④ Process only specific segments / temporarily change runid or resolution
python3 $S/scripts/make_ad.py <project.json> --steps fetch,verify --seg 3 --runid v24 --res 720P

Each step checks exit code + file existence/size. If criteria are not met, it reports the FAIL reason directly and lists it — does not silently continue.

Project Configuration (templates/ad_project.hsq1.json is a real, directly runnable example — modify it to create a new video):

FieldDescription
modelEngine (Use wan3.0-video for product shots; default wan2.7-r2v only suitable for pure character scenes)
productPoints to a key in product_profiles.json — Shared between image generation and video generation
refsReference image order = Image 1..Image N, must correspond one-to-one with image descriptions in the product lock template; local images are automatically scp'd to a public URL
segmentsStoryboard text for each segment (Only write "what to shoot"; product lock is automatically prepended by make_ad)
assembly.voVoiceover (edge-tts); timeline = subtitle start, naturally aligned
assembly.subs/chips/spec/complianceSubtitles / stylized text / parameter bars / AI-generated content marking
assembly.cut1515s version: which segments to take, where to cut the voiceover, new time window for each overlay on the 15s timeline
deliverDelivery directory + filename (default ~/Desktop/电商素材/<product>/视频/)

Product Lock (shared across all products, do not create new ones): templates/product_lock_head.txt + video_lock field in the product profile. "Proportion ≤ 1/6 of frame height, measure using head diameter as ruler (1.9x), bottom end is a flat circular surface, line inserted along the axis, never look directly at the lens…" These hard-learned phrases are all there — Do not repeat in storyboards, and do not hand-write new templates.

Empirical Evidence (2026-09-18): Running ad_project.hsq1.json with --steps assemble,deliver produces final videos whose md5 checksums perfectly match the manual delivery version from the night of 2026-09-17 (30s 86f3da96…, 15s 0b67686b…); --steps plan,submit,fetch,verify single-segment smoke test successfully completed (submission → polling → download → frame extraction verification).

Component List (scripts/ 26 items, grouped by purpose)

StepScriptPurpose
Mainmake_ad.pyOne-command video generation (plan / submit / fetch / verify / assemble / deliver)
Generationvideoretalk_submit.py / probe_model_availability.py / compare_video_models.pyLip-sync submission / Model availability probe (successful submission ≠ availability) / Cross-model comparison for same prompt
Verification · Productproduct_fidelity_zoom.py / scale_consistency_check.py / scale_panel_check.py / measure_product_truth.py / check_reference_image.py / locate_reported_frame.py / audit_shots.py2x product zoom / Cross-shot size consistency / Truth + output frame reading panel / Real product aspect ratio ground truth / Reference image health check / User screenshot timestamp lookup / Per-shot audit
Verification · Audiostt_dialogue_check.py / voice_consistency_check.py / voice_timbre_similarity.py / tts_fit.pySpeech-to-text reverse check of lines (subtitle timeline source) / Voice change detection / Voice timbre locking verification / TTS speed adjustment for shot alignment
Verification · Visualverify_motion.pyMotion detection — determines if the final video is a "static image collage"
Post-productionassemble_r2v_overlays.py / build_vo30.py / overlay_anim.py / replace_segment_audio.py / gen_ass.pyAssembly (supports vo full voiceover track) / Voiceover track / Stylized text animation / Replace audio within segment / ASS stylized text
Zero Distortion Pathbuild_official_product_insert.py / compose_official_material_ad.py / cutout_product.py / make_comp_cards.pyOfficial pixel-perfect product insert (top choice for product shots, 0 cost) / Official material dynamic effect video / Official image cutout to transparent PNG / Composite cards
Charactersgen_character_candidates.pyMale lead candidate generation (produces 3-4 collages at once for user to choose)

Path Decision (Most Important · Hard-Earned Lessons; Rewritten 2026-09-19)

PathEffectWhen to Use
① Product Lock + wan3.0-video ✅ Current Main PathWith same prompt and same 5 reference images, product is correct on first try, no degradation across entire segment; realistic human life scenarios. Reference images = macro mesh shot / front full body / real hand scale / bottom end face / male leadMost promotional videos (have official product images, require "product looks like real device")
② Official Pixel-Perfect Product Insert (0 API Cost) ✅ Ultimate Product Fidelity SolutionProduct 100% equals real device (official pixel-perfect, no AI involved)When AI consistently fails to draw the product accurately; or when the video needs viewers to instantly recognize the product
③ Text-to-Video (wan2.6-t2v, lines via input.audio_url)Real people + scene + Chinese lip-sync, but product changes slightly in each segmentNo official materials, or purely character-based story
④ Image Stitching (clipforge / MoneyPrinterTurbo)❌ User verdict = "image collage, lacks realism, far from sample quality"Only for draft placeholders, not acceptable as final delivery
⑤ Digital Human Lip-Sync (wan2.2-s2v / EMO / videoretalk)Single person voiceover, weak environmental realismOnly need voiceover narration

Recommended Combination: Character/life scene shots → ①; Product appearance/close-up shots → ②. Final video structure = "Hook 5s (AI) + Official product insert 3s (pixel-perfect) + Main content 7s (AI)" = 15s; Script scripts/build_official_product_insert.py (crops official image to create 3s vertical product insert, subtitles/stylized text offset by +3s).

Ironclad Rules:

  • ❌ wan2.7-r2v for product shots is retired (product appears small/blurry/easily washed out to silver-white = root cause of repeated rework); use only for character scenes
  • Generation is always at 720P (native 1080P generation = re-rolling, not an upgrade; in practice, product deformation/loss of gold/details no sharper) → Final video is upscaled to 1080 during assembly
  • wan3.0-video treats reference image 1 as the first frame (control the opening by changing image 1)
  • If the user's "sample standard" shows a real person dialogue scene, do not assert it "can only be shot live" (user explicitly corrected this: their sample itself was AI-generated)
  • Image stitching (④) can never match the realism of ①; do not pass it off as an AI video deliverable; directly use ①/② with the user's own lines and voiceover

Backup Route Recipe: wan2.6-t2v (Real Person Story · Product May Vary)

Complete parameters, JSON, billing, voiceover hosting, and verification commands are in references/scene-drama-and-candid-style.md.

Quick Highlights:

  • POST https://dashscope.aliyuncs.com/api/v1/services/aigc/video-generation/video-synthesis model=wan2.6-t2v, header X-DashScope-Async: enable
  • parameters: size:"720*1280" (9:16) or "1080*1920", duration:5|10, shot_type:"multi", prompt_extend:true
  • input.audio_url: Pass our own written lines as audio (3~30s mp3/wav public URL) → Model performs lip-sync based on the audio; If not provided, the model auto-generates sound effects and makes up its own lines (uncontrollable, selling points will drift)
  • Pricing: 720P 0.6 CNY/second, 1080P 1 CNY/second; account has 50 seconds free quota → First 30-second video can be generated at 0 cost using quota
  • Voiceover timbre: edge-tts zh-CN-XiaoyiNeural (young female) / zh-CN-XiaoxiaoNeural (older female) / zh-CN-YunxiNeural (male)
  • In the prompt, use 1st shot [0-4 seconds]…2nd shot [4-10 seconds]… to describe multiple shots; add "No text, subtitles, or watermarks should appear in the video" at the end

Standard Production Workflow

  1. Script: 3 segments × 10 seconds (or 6 segments × 5 seconds): Story hook → Product introduction + real selling points → Conclusion + CTA
  2. Voiceover: Generate mp3 for each segment using edge-tts → scp yunvela:/tmp/ → sudo mv to /opt/yunvela-site/dh/ → curl -o /dev/null -w '%{http_code}' https://yunvela.com/dh/segN.mp3 must return 200
  3. Generation: Submit segments one by one (can be parallel), poll GET /api/v1/tasks/<id>, curl -sL --retry 3 --retry-all-errors to download
  4. Post-production: Stitch + stylized text (white text with black outline + yellow keyword box) + emoji stickers + BGM → Final 9:16 video
  5. Verification (cannot be skipped) : ffprobe for duration/resolution/audio track presence + volumedetect to confirm audio is present + extract frames for visual inspection + STT reverse check of lines (~/.hermes/scripts/voice_stt_local.py) to confirm the spoken words match our script
  6. Delivery: Sending large videos via WeChat gets throttled by iLink → ~/.hermes/scripts/feishu_send_media.py <video> "description"

User Preferences (Confirmed in This Project)

  • 🔴 Do not entrust product close-ups to AI — use official product images for dynamic effects (empirical conclusion from 2026-09-15): In AI-generated character story segments, the product "varies slightly" in each segment (our test: under the same prompt, one segment shows a constant-diameter cylinder, another becomes a tapered rod with a mesh head, and even repeated logos appear). Correct approach: Hand character scenes to AI, for product close-ups, use official product images with a slow Ken Burns zoom-in (scale=1080:-1,pad=1080:1920:(w-h)/2:(w-h)/2:white,zoompan=z='min(zoom+0.0007,1.10)':d=<number of frames>), and attach the original segment's voiceover to the official image dynamic effect (no sentence break). This makes the product 100% consistent with the real device, satisfying the client's requirement that "the product must be realistically integrated."
  • 🔴 Two reference images must be in the "same orientation and form": Both product reference images for r2v must be taken from a vertical perspective (our test: using a "bottom end cross-section view" as the second reference caused the model to average between slender and fat, producing a squat cup-shaped product). Do not mix top-down/upward/cross-section images as references. Reference image process: Crop to object bounding box → Paste onto a white square canvas, with the object occupying about 60% of the canvas height.
  • ⚠️ AI may burn in its own subtitles: Even with prompt saying "no text", wan2.6-r2v-flash often adds built-in bottom subtitles. Therefore, place our stylized text at the top (Alignment=8 + large MarginV), and do not overlay another layer at the bottom → otherwise double subtitles will overlap.
  • Determine dimensions based on actual measurements, not visual estimation (lesson from 2026-09-15): The prompt's "height:diameter" must be calculated from official dimensions (HSQ1 = 74×39mm → height:width 1.93:1 slender cylinder; I once visually estimated and wrote "squat cup shape 1.2:1", and the client judged the result as "still inaccurate"). Example wording: The body is a slender cylinder, approximately 74mm tall and 39mm in diameter (height about 1.9 times the diameter), with a black matte satin finish…
  • Ironclad rule for messaging (user correction on 2026-09-15): Lines/voiceover/stylized text must only cite official promotional material verbatim (folders, main image text). ❌ The "selling points" field in the attribute table ≠ official messaging (internal descriptions); ❌ Claims heard from reference samples/competitor videos must not be used as messaging (HSQ1 test: "egg-sized" is banned); 🟡 Size-related comparative terms (e.g., "thumb height") can only be used for visual proportion/comparison objects, not in messaging. → Before writing lines, first extract official messaging using product-spec-extraction; do not write selling points from memory.
  • Results first (user decision on 2026-09-15): Video production prioritizes results; do not compromise visual quality for cost savings.
  • If there is output, must proactively report progress; for long workflows, provide segmented progress updates; do not wait for follow-up.
  • During the design phase, write all real selling points; do not delete them due to "lack of evidence" (proof is the placement team's responsibility); for restricted terms/efficacy claims, only issue warnings without blocking.
  • Cost-sensitive: Use free quotas/low-cost options to validate before scaling up; state unit price before spending money.

Related Skills

  • clipforge-video-pipeline: Complete recipe for image stitching solutions (clipforge / MoneyPrinterTurbo) = Path ④, includes environment installation pitfalls, shot-by-shot image placement, local asset mode, and the 0-cost "real product dynamic clip library."

📚 Knowledge Base Index (8 Documents)

When to ReadDocument
Product looks wrong / deformed / discolored / size and proportion / meshreferences/product-fidelity.md
Choosing an engine / model fields / pricingreferences/engines-and-models.md
Resolution / billing / arrears troubleshooting / pre-checksreferences/resolution-cost-and-arrearage.md
Main pipeline / assembly / delivery verificationreferences/pipeline-assembly-and-qc.md
Completion rate / hook variations / performance metricsreferences/hooks-and-ad-metrics.md
Voiceover / subtitles / timbrereferences/voice-and-subtitles.md
Portrait rights / compliance red lines / reference image standardsreferences/compliance-portrait-and-references.md
Scene and story videos / realistic shooting techniques / storyboard wordingreferences/scene-drama-and-candid-style.md

Individual skills in this repo

This repo contains 1 individual skill — each has its own dedicated page.

相关技能