Build one correct creative.
Then build the next test pack.

B2C means ads for homeowners. The output is a short vertical video inviting a homeowner to request an inspection or estimate. A pack is a named set of videos testing one clear question, plus the files and evidence needed to reproduce them.

Definition of done: one playable 1080 × 1920 control, correct words and voice, readable top banner, captions that follow the words without covering the repair, complete playback review, and a manifest mapping every input to its output.

ffmpeg -version
ffprobe -version
py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install Pillow
.\.venv\Scripts\python.exe -c "from PIL import Image; print('Pillow ready')"

Extract the kit into C:\StPierre\creative-kit. Run commands from that folder in PowerShell. Keep inputs in inputs, silent edits in work, finals in outputs. Create those folders first. Commands below are for the builder’s PC, not Akash’s Mac.

Ready check: open one source MP4, play the supplied Audio B, run both FFmpeg tools and import Pillow. Then make the practice render in E. Windows-native execution has not been tested on the builder’s hardware; the kit’s render path is smoke-tested on Akash’s machine with an explicit font.

Drive links require the builder’s existing Google account access. A link does not grant access. If access is denied, Akash must share the specific folder with the builder; this task has not changed any permissions.

1. Open the B2C creative scoreboard ↗ or obtain a current export. Record the date window, timezone, attribution setting, market and destination. Do not use the September 9 snapshot as a live campaign status.
2. For each ad, join its spend to its actual creative and downstream homeowner record. Record Meta leads, conversations and qualified inquiries separately. Confirm the definition and attribution of “qualified” with the operator.
3. Calculate cost per qualified inquiry = spend ÷ qualified inquiries. A zero denominator means “no qualified result,” not $0. Rank within a comparable market/window/offer; retain small-sample caveats. Use cheaper views or likes only to diagnose, never as the paid winner criterion.
4. Watch each control from start to finish. Log first-frame visual, repair action, top text, voice, spoken script, caption layout, length, offer and destination in audit.csv. Match by file/frame evidence, not a similar filename.
5. Find what is missing: a better opening, an underrepresented repair mechanism, weak narration or unreadable captions. Do not change footage, voice and offer together and then attribute the outcome to one of them.
6. ★ First exercise: reproduce the existing foundation Audio B look from supplied assets. Then prepare the three existing B variants shown below as a footage test. No generation is needed to compare those already-built files.
7. Agree the market, active receiving client, test budget, review date and stop rule before activation. Draft renders do not authorize spend. Refresh qualified outcomes before final selection. The broader September 8 “20 hooks × 2 formats” model was unresolved for B2C video; it is not a mandatory 40-video batch.

Reusable script structure: identify the service/problem → show or explain the fix → address the relevant concern → invite the next action. Foundation repair uses an inspection; concrete leveling often uses an estimate. Concrete flatwork sells replacement, so it must not inherit “without replacement” copy.

Top text vs captions vs ad copy: TOS here means the persistent text-on-screen hook (“Foundation Problems?”). Burned captions are the spoken words inside the video. Meta primary text and headline sit outside the video. Terms of service are a separate meaning of TOS and are not the copy bank.

Real footage: preserve the original, work from a clean copy

1. Start with an approved clean clip or an intact proven master. Store source URL, file hash, clip ID, owner/permission evidence and approved status. “Clean,” “approved historically” and “licensed for this use” are distinct checks.
2. For long raw footage, sample a contact sheet to locate useful work, then inspect the entire chosen range. For competitor archives, deduplicate before reviewing. Keep the source file untouched.
3. Choose a short continuous repair action. Trim intros, logos and end cards from material you are authorized to edit. Remove source audio. If captions/branding cover the mechanism, use a different clip or generate the missing shot; a blur is not a substitute for rights or visible evidence.
4. Prepare a clean silent base and scan the whole range for surviving text/logos. The old automatic OCR tool uses Apple Vision and is Mac-only. On Windows, review extracted frames plus full playback; do not report an OCR failure as zero detections.
5. Foundation visual standard from August 13: show piers, jacks, drilling or crews doing repair work. Symptom-only shots were cut from that repair pack. Later symptom-to-fix AI concepts are a separately labeled experiment, not an undisclosed replacement of the control.

ffmpeg -i "inputs\source.mp4" -ss 2 -t 6 -an -c:v libx264 -crf 18 "work\trimmed-silent.mp4"
ffmpeg -i "work\trimmed-silent.mp4" -vf "fps=2,scale=360:-2" "work\scan-%04d.jpg"

Higgsfield: generate only the missing visual

Read D1 for the full realism workflow, Genjutsu modes and copyable prompts →

The August 26 process reviewed real frames to describe shot composition and mechanics. Competitor frames stayed analysis-only. Generate a new scene from written descriptions; use your own generated stills as references for matched before/after shots.

1. Write the shot brief: visible symptom, repair mechanism, fixed camera, stable in-frame reference, amount of motion, and final result. Example: “Locked phone camera at ground level. A hydraulic ram pushes a steel foundation pier downward beside an excavated footing. Worker checks the connection. Clearly visible small mechanical travel. Fixed brick line stays stable. Realistic materials, no text, no logos.”
2. For matched before/after: generate an original “before” still; use that generated still to create an “after” with the same framing. Animate your generated frames. Do not manufacture a client testimonial or present an AI repair as photographed proof of a completed job.
3. Use the Higgsfield browser workflow or CLI. Discover current model settings. Current skill default for raw motion is Seedance 2.0; earlier outputs are not all the same model. Save actual model/job/settings alongside every result.
4. Generate one shot first. Inspect hands, tools, material continuity and actual visible movement. Reject melting workers, impossible mechanics, flickering text and unexplained jumps between states. A repair-process shot must communicate the intended work; a restrained finished-room pan can serve as ambient B-roll but does not prove an active repair. Review the prototype before scaling a batch.
5. Download accepted output, remove generated audio, register it as AI/illustrative, then assemble with the same locked audio and overlays as the real control. Record any credits consumed. Model-generated audio does not replace Eric in a locked voice test.

npm install -g @higgsfield/cli
higgsfield auth login
higgsfield account status
higgsfield model list --json
higgsfield model get seedance_2_0 --json
higgsfield generate create seedance_2_0 --prompt "YOUR APPROVED SHOT BRIEF" --start-image "inputs\approved-start.png" --aspect_ratio 9:16 --duration 5 --wait

Generation consumes account credits. The final command is an example to run after the shot and allowance are agreed. Correct login syntax is higgsfield auth login; the old installer’s higgsfield login message is stale. Validate parameters against the current model before submitting. Official Higgsfield CLI instructions ↗

Voicebox: Eric is built in, the treated take is not

1. Install/open Voicebox on Windows. Select the Kokoro engine, English male Eric / am_eric. Create a profile called CTR-Eric. That label is our convention; its new local profile ID will differ from Akash’s. Do not select a different engine simply because it has a voice with the same name.
2. Paste the exact approved script. Explicitly choose Kokoro when generating; do not trust the default engine. Export a WAV. Save script, engine, preset, app version and export together.
3. For the first foundation pack, use supplied option-a/b/c.wav directly. These are the treated takes already used in the renders. Do not speed them up a second time or regenerate them and call them the same control.
4. For a new script, first apply 1.06× tempo to the raw take. Then manually shorten excessive sentence gaps against the waveform, keeping a separate raw and treated file. Historical longer reads used roughly 0.08 seconds between sentences and a protected 0.45-second beat before the ask. Short Audio A/B/C have their own locked timing; those takes override the generic pause recipe.
5. Listen to the entire treated WAV. Check pronunciation, breaths, clipping, chopped syllables and the final ask. No music in the current voiced control. Other voice auditions were listening tests; no paid music-track winner was found.
6. Only after audio is final, create the word clock. Whisper provides timing, while the approved script supplies the words. Fix recognition mistakes and alignment manually before drawing captions.

ffmpeg -i "inputs\raw-voice.wav" -af "atempo=1.06" "work\tempo-voice.wav"
ffmpeg -i "work\treated-voice.wav" -ar 16000 -ac 1 "work\alignment.wav"
whisper-cli -m "models\ggml-base.en.bin" -f "work\alignment.wav" -osrt -ml 1 -of "work\words"

The Whisper model must be downloaded from its official project; the example filename is a local model path. -ml 1 is the prior workflow’s short-segment timing route: inspect that the resulting SRT contains one spoken word per nonblank entry. A matching word count alone does not prove correct alignment. See Whisper instructions ↗ and Voicebox source / downloads ↗.

1. Build the silent edit. In a timeline editor, create a 1080 × 1920, 30 fps composition. Import chosen shots. Set the duration to the treated voice. Preserve the actual sequence of the repair. Use a deliberate crop only when it keeps the mechanism visible; for landscape footage, preserve the full footage band on background padding.
2. Match length without hiding bad motion. Trim unused lead-in/tail, use multiple compatible shots, or moderately retime the action. Past sessions rejected obvious loops and overly stretched footage with melting workers. Keep speeds within roughly 0.5–4× and inspect the result. Do not silently loop a repair or slow it to 0.3× to fill a longer read.
3. Export the silent base. No narration, banner or captions yet. Ensure it matches the treated audio duration within 0.15 seconds. Name it work/base.mp4. The portable tool deliberately refuses the wrong size/duration instead of deciding a destructive crop or loop for you.
4. Write top text. Pick a short line from the bank matching the visible problem. Foundation control: “Foundation Problems?” A white rounded banner with dark bold type remains through the clip. Do not invent prices, guarantees, discounts or an offer the receiving client cannot honor.
5. Author caption cards. Split approved words into 2–3 word groups at natural phrase breaks; 1-word cards are allowed for long terms. Commas/full stops generally become ellipses; preserve question/exclamation marks. Supplied cards already match A/B/C. For new audio, compare each word with its timing, not just the total count.
6. Render. The tool adds the treated WAV, draws the top banner, and burns white text on red cards with the active word’s text yellow (#FFD400). Font starts at 5.7% of frame height and shrinks long cards to fit 94% width. It uses the corrected end-exclusive caption timing to prevent two cards occupying a boundary frame.
7. Place captions per clip. Caption-y is a fraction of full height: 0.68 is 68% down. Choose dead space away from the actual drill, jack or crack. For padded landscape video, put the card in the padding. Past FR edits put cards over blur seams where that also preserved action. Never hide the mechanism just to hide a seam.
8. Save proof. Watch the exported file at full speed with headphones, then muted. Frame-step across each card transition and cuts, including the first/last frame. Save QA and the manifest; a successful encoder is not a visual pass.

.\.venv\Scripts\python.exe tools\build_video.py --video "work\base.mp4" --audio "audio\option-b.wav" --cards "audio\option-b.cards.json" --words "audio\option-b.words.srt" --banner "Foundation Problems?" --font "C:\Windows\Fonts\arialbd.ttf" --caption-y 0.68 --out "outputs\FR-practice-b.mp4"

The renderer adds a .build.json receipt and performs a full decode. It makes no network calls or live changes. You must still review timing, crop, offer, source marks and realism. It refuses to overwrite an existing output; use a new revision name when fixing a render.

Make the silent base without a timeline editor

For a single continuous approved shot, the included helper preserves the full frame, pads to portrait and retimes it to the supplied audio. It does not loop or invent frames. Save the clean clip as inputs/source.mp4; create work and outputs folders. This is a practice route; inspect movement and padding before using it in a paid test.

.\.venv\Scripts\python.exe tools\prepare_base.py --video "inputs\source.mp4" --audio "audio\option-b.wav" --out "work\base.mp4"

Then run the caption/banner command above. The helper stops if matching the audio would require speed outside 0.5–4×. Select another range or assemble more shots rather than forcing an unnatural result. White padding is the portable default; an existing source with blur padding keeps that source appearance.

When a manual editor is useful

Use it to select shots, trim, change speed, preserve landscape framing, adjust audio gaps and inspect transitions. Use our renderer for the exact small cards and word highlight. Existing finished masters already have their captions; reuse intact as controls. Re-voicing a master with burned text requires matching the original words and timing or rebuilding from clean footage—otherwise it becomes contradictory.

Next actions, with one owner per handoff

First pack recommendation: start with the existing real / lift / symptom Audio B foundation variants for Pittsburgh, if current lead quality and client availability support it. Dallas’s lower observed-cost Audio A is a separate market starting point. These are small-sample suggestions, not declared winners. Do not generate another large matrix before learning from the existing files.

Beyond foundation repair: the same audit → source → audio → overlay → QA loop serves concrete leveling, crawl space, waterproofing and flatwork. Change the symptom, repair mechanism and offer to the niche. Their library links and copybank rows are above; foundation examples below teach the mechanics.

• Live Meta data was pulled September 9 for August 25–September 8, 2026 inclusive, from the Foundation Repair account. One “lead” action is counted, not duplicate grouped aliases.
• Historical qualified counts are from the saved August 25 source selection. Newer qualified status and current client availability were not refreshed because database queries timed out.
• These are observed campaign results, not a randomized study. Do not compare Dallas and Pittsburgh as if only creative changed.
• Local manifests sometimes still say “pending upload.” Actual Meta spend proves that the nine Pittsburgh variants and three Dallas variants received delivery; stale local labels do not override that evidence.
• The embedded files are full local source renders with H.264 video and AAC audio. Main audio players are exact extracts from the corresponding real-footage render.