← سند dashboard

سند UGC reel guidelines (LIVING DOC)

Owner: Bandar. Maintained by the Video Editor (سند) with the سند Lead. Created 30 Oct 2026, 1:30 PM (Riyadh) after reel 1 v6 was locked. Sources: Bandar's messages relayed by the Lead (ids in brackets), the Higgsfield usage log (/workspace/higgsfield/usage.log), reel-1/test/NOTES.md, PRODUCTION-PLAN.md, STANDING-RULES.md, brand-face/identities/README.md, the higgsfield-creative skill and the reel 1 work files.

0. UPDATE PROTOCOL (Bandar, pt537u: "update this document immediately after I approve each new job, to self-improve the video-making process; align with the video editor")

  1. The moment Bandar approves (locks) any reel, card, template or fix: copy the approved file to …/APPROVED.mp4 (read-only), then in the same turn add (a) any new preference to section 2, (b) the fix or technique that worked and any failure to avoid, (c) the credits spent to section 5, (d) a line to the changelog (section 7).
  2. Never edit a preference's wording; if Bandar changes his mind, add a new row that supersedes it and mark the old one SUPERSEDED.
  3. The Video Editor builds from this doc; the Lead updates it with new Bandar messages. Both read it before the next reel.
  4. Never overwrite an APPROVED file. New versions get a new name (-vN-candidate).

1. Production pipeline that worked (reel 1)

  1. Script approved first (Arabic, Copy Humanizer-cleared, Bandar's exact wording wins). Nothing is generated before approval of the script, then of the stills.
  2. Voice: ElevenLabs eleven_v4_turbo only, one MP3 per line (L1…L8), never Bandar's clone. Reel 1 used the stock library voice «NoraS_F» for Jawaher (voice_id FZeLZd39ejvLgzR2gY0t; stability 0.5 / similarity 0.8 / style 0.2, language ar). Mix all lines to one vo.wav; keep a timing.json with each line's start.
  3. Stills: from the identity hero + refs (nano_banana_2_1, 2k, about 2 cr each), framed half-body/chest-up, 9:16, same still for every model under test.
  4. Motion: Kling 3.0 std 720p image-to-video, silent, neutral mouth, negative prompts; talking-head only where a face speaks, everything else is voice-over on our own UI cards (0 credits, built in HyperFrames/GSAP).
  5. Lip-sync: Sync Lipsync 3 (sync_so) on the talking segments with the untrimmed ElevenLabs audio for that segment.
  6. Assembly: HyperFrames 0.8.114 (build.py -> index.html, Tajawal, captions + cards in GSAP, vo.wav), render --workers 2 --crf 24, then ffmpeg -c copy -movflags +faststart. 1080×1920, 30 fps. Deterministic: unchanged parts re-render pixel-identical (SSIM 1.000), which is how "keep everything else identical" is proven.
  7. QA (section 3) before showing Bandar. Save as -vN-candidate, report path + credits + contact sheet.

2. Bandar's preferences (every one logged, with what worked and what to avoid)

# Date / id Preference (Bandar) Fix / technique that worked Failure to avoid
P1 30 Oct, pt525u Full consistency between halves of the reel: same face, wardrobe, light, framing and the same level of expression. "First half is perfect, fix the second part after «أو», make it consistent." Keep the approved first half frame-identical; regenerate only the weak part, from the last clean frame before the problem as Kling start image (same model, same Jawaher refs), so the seam is invisible (4-frame blend at a pause). Re-doing the whole clip (drifts the approved half); a second half that is more/less smiley or expressive than the first.
P2 30 Oct 12:11–12:46, pt525u, pt530u Synced lips, always. Voice and mouth must match; visible articulation on every word. Sync Lipsync 3 with the full, untrimmed VO for the segment, sync_mode: silence, a video 0.4–0.5 s longer than the audio. Neutral-mouth Kling base so Sync has room to articulate. Trimmed audio: v3's L6_trim.mp3 clipped the tail of «لك» so the mouth closed 0.3 s early. A base clip with a fixed smile: Sync keeps the smile and barely moves the lips (v5 "same issue").
P3 30 Oct 12:46, pt530u Check for closed-mouth frames during speech. Sample the final reel every 0.2 s; MediaPipe FaceLandmarker lip gap (L13/L14 over face height) vs speech energy (log-RMS of the VO); require mouth open/moving in the loud frames, then look at a 0.2 s contact sheet by eye (section 3). Trusting a timing/lag check alone: it found no drift in v5 while the mouth was shut. Trusting a single metric (it is noisy).
P4 30 Oct 12:46 Retiming or holds can freeze the mouth, so don't rely on them. If the mouth is wrong, regenerate that segment (Kling + Sync), do not retime. v4/v5: holding a frame / mouth-patch crossfade to delay the smile produced a frozen closed-mouth smile while «لك» was still audible.
P5 30 Oct, standing No odd body movements; natural hand timing. One small presenting palm gesture that rises on the beat of the line and lowers softly before the last word; no jitter, no melting hands, no extra fingers. Prompt for exactly one timed gesture ("hand rises after ~1 s, lowers by ~4.5 s"), negative prompts, inspect a 2 fps contact sheet for the hands. In v6 the gesture is different from the approved take (later, gentler) and was accepted. Letting the model improvise gestures; a hand held up while the face goes neutral.
P6 30 Oct 12:40, pt529u The smile comes only after the final word (v4: smile started before «لك» finished). Say it in the prompt: neutral warm expression throughout, "NO grin, NO teeth-smile while talking", gentle closed-lip smile only in the last half second; then confirm with the blendshape mouthSmile curve against the last word's end. A prompt that mentions the smile early ("smile appears at ~5 s") made Kling grin through the whole last phrase (v6 attempt 1, discarded). Even in v6 the talking-smile ramps from about 7.1 s of the voice line (25.4 s of the reel), before «لك» ends. Known imperfection, accepted by Bandar; try harder next time.
P7 30 Oct, standing Captions timed to the words (Tajawal Bold, white on #2121F5 pill, 2–4 words, safe area). Derive caption starts from the word timings of the VO (v2/words.json), not from the line start. Re-check after every re-cut. Caption «أو تشغّل التقديم التلقائي» appeared about 0.9 s after «أو» (fixed).
P8 30 Oct Refine the closest model rather than redo; test on several models first (Kling, Seedance, Wan). One shared still + prompt + VO across Kling 3.0, Seedance 2.0 Fast, Wan 3.0; Bandar picked Kling; then v2 → v6 refine Kling only, reusing good parts (S3 hands reused from v1). Restarting from scratch each round (spends credits, loses the approved half). Veo 3.1 needs the Plus plan and Veo 3.1 Lite returned rate_limit: not usable on Starter.
P9 30 Oct, STANDING-RULES 9:16, 1080×1920, about 42 s or less. Render at 1080×1920/30 fps; check ffprobe. Any other ratio or a Kling clip exported without the 716×1284 crop being scaled to cover.
P10 30 Oct, STANDING-RULES Link in the bio; never a URL on screen, in captions or in end cards. Pill «الرابط في البايو ↑» on the save/share card and the end card variant -bio. Writing «sanad.wa3e.ai» anywhere (reel 11's spoken URL and reels 3/10/14 end cards were removed).
P11 30 Oct, STANDING-RULES; redesign 1:30 PM Ask viewers to save and share, before the end card. Current on-screen line (Bandar 1:30 PM): «احفظ الفيديو وشاركه مع شخص يحتاج وظيفة» (his "الفديو/وشاركك" read as typos). Template brand-templates/save-share-card/ (blue panel, bookmark + paper-plane buttons, single line, bio pill). Two separate text lines and two plain circles (old reel 1 card, replaced in v7). Spoken L7 and on-screen wording must match in the next reel; reel 1's VO still has the older wording.
P12 30 Oct 12:13, 12:25, 12:37, 1:26 PM Signature end card on every سند video/image: logo, the canonical slogan, «نقبل الدفع عبر تابي وتمارا», real Tabby and Tamara logos in real colours on white chips, 9:16, no URL. Slogan is now «خلّ ورا ظهرك «سند»» (replaces «خل في ظهرك سند», 30 Oct 1:26 PM). brand-endcard/ v3 (PNG + MP4 + bio variant + layers + source). Logos unmodified from the official kits (Tabby media kit TabbyBadge.svg; Tamara_Brand_Assets "Wordmark Gradient"). White single-colour logos (v1, rejected); recoloured/outlined logos. Do NOT publish "we accept Tabby" until Tabby is live (no API keys yet). Check the inner «» around سند with Bandar on first use; reel 1 v6 (locked) keeps the old slogan.
P13 30 Oct, standing سند palette only: #2121F5, #FAF8EA, white. Never neon green. Only exception: real third-party logos (Tabby, Tamara). Tajawal, blue pills, cream cards. Neon accents, gradients outside logos.
P14 30 Oct, decided Approved apply wording: «سند ما يقدّم إلا بموافقتك على كل وظيفة، أو تشغّل التقديم التلقائي ويقدّم عنك على الأنسب لك.» (without the leading «سند» after a سند line). Used verbatim in L6 of reel 1; caption split «بموافقتك على كل وظيفة» / «أو تشغّل التقديم التلقائي». "Stop or edit through the team" lines; any claim that سند applies without approval or opt-in.
P15 29–30 Oct Brand faces: Faisal (فيصل), Hessa (حصة), Jawaher (جواهر), Abdulaziz (عبدالعزيز, added 31 Oct 2026, P24; voice TBD, no Soul ID yet), fictional AI characters, labelled AI-generated, never real clients/staff/testimonials. Voices: ElevenLabs eleven_v4_turbo only; Faisal = Firas_M; Hessa and Jawaher final voices to be picked by Bandar (reel 1 used NoraS_F for Jawaher). Identity files in brand-face/identities/<name>/ (IDENTITY.md, refs, Soul IDs: Hessa 8bd08845-…, Jawaher 034094bb-…, Faisal 30484346-…). Start talking shots from half-body or chest-up stills; re-check glasses/agal each clip. Wide hero frames as the Kling start (face too small = soft); two Soul faces in one shot (not possible, alternate shots).
P16 30 Oct 1:26 PM Progress rings: the number centred inside the ring, label «آخر المستجدات». (Dashboard/progress cards; reels 2, 14 and any status card.) Build in HyperFrames with the number absolutely centred (flex centre, tabular numerals), label below the ring in #2121F5 Tajawal. Off-centre numbers; other label wording. (Not yet used in an approved reel: confirm the look on first use, then update this row.)
P17 30 Oct, standing Scripts approved before any generation; Bandar's exact Arabic replaces the closest existing line (section 0b of PRODUCTION-PLAN). Approve scripts → approve stills → generate → QA → lock. Generating (and spending credits) on unapproved copy.
P18 30 Oct, standing Higgsfield through Mira's Studio and the shared credit rules. One shared pool on the Starter plan; check the balance before any job over 20 credits; append a line to /workspace/higgsfield/usage.log after every session; download every output; no top-ups, no plan changes (Bandar/Emily decide); Starter cannot use Seedance 2.5, Seedance 2.0 full, Veo, start+end frame control, Speak/UGC builder. Max 2 videos and 4 images at once. Spending beyond the stated cap without asking; subscribing or changing plan from any checkout page.
P19 30 Oct, standing Claims only about real features; no invented stats/price/promo, «5,000+ employers» unverified and not used; matching/dashboard screens labelled «عرض توضيحي · بيانات تجريبية»; do not post matching-dependent reels until matching is live. Cards built by us, always labelled. Showing mock results as real.
P20 30 Oct 12:46 Report format: path, credits, and a contact sheet PNG of the part that changed. reel-1/sanad-reel1-v6-second-half-contact-sheet.png (0.2 s tiles with time, mouth opening, VOICE flag). Reporting only metrics.
P21 30 Oct 2:00 PM pt548u; widened 3:06 PM pt556u to three platforms ⏸ PAUSED 31 Oct 2026 10:18 AM by P25 (cuts only on "repurpose"). Every piece of content Bandar approves (video OR image) is repurposed for THREE platforms: X, Instagram and TikTok, each built for its own safe margins and format (section 3b), and ALL files are uploaded to Google Drive (folder Sanad Reels/Reel N, or Sanad Content/<item> for images) for him to download. Naming: sanad-<item>-<Platform>-vV-YYYYMMDD.<ext> (platform, version number, date), e.g. sanad-reel1-X-v7-20261030.mp4. Send Bandar the Drive folder link plus each file link. Video: 1080×1920 9:16 for TikTok and Instagram; X also 1080×1920 (X plays vertical video), keep text inside the X safe box and the essential content inside the centre 4:5 area, since X crops vertical video to 4:5 in the timeline preview. Images: Instagram 1080×1350 (4:5), TikTok 1080×1920, X 1600×900 (16:9) or 1080×1350. Audio is never changed. Re-render the overlays per platform (reel-1/platform-v7/make.py <tiktok|instagram|x>), copy the audio from the approved file, and draw a check image with the safe zones (platform-v7/check.py). 0 credits. Do this immediately after approval, without asking. One master posted to every app; text under any platform's UI; forgetting X or the Drive upload.
P22 31 Oct 2026 9:03 AM, pt572u ("I like D, lock it in") Lip-sync engine preference: HeyGen Avatar IV (D) over Kling 3.0 + Sync Lipsync 3 (A), OmniHuman 1.5 (B) and Kling AI Avatar v2 Pro (C). Bandar prefers it because it is more expressive (head motion, wide articulation). The constant grin (mean smile 0.63, max 0.95) was noted and accepted; it supersedes the "no grin while talking" wish of P6 for HeyGen talking-head pilots only (P6 still applies to Kling/Sync reels). Pilot clip (reel 13, Jawaher, 10 s line line.mp3, one front still): HeyGen v3 API POST /v3/videos type:image + audio_asset_id (own ElevenLabs MP3, no HeyGen TTS), 1080p 9:16, default expressiveness, ≈4 min render. qa_lip: mouth open 77% of loud frames, moving in 59% of 0.5 s windows (A 51%/22%, B 67%/48%, C 77%/41%). Costs (same 10 s line): D $0.40 (wallet 5.00 → 4.60, ≈$0.04/s); A 53.5 Higgsfield credits; B $1.62 (10.109 s × $0.16/s, fal); C $1.17 (10.133 s × $0.115/s, fal; B and C computed from billed duration, the fal key cannot read the balance). Approved file reel-13/sanad-reel13-pilot-heygen-D-APPROVED.mp4; compare grid reel-13/pilot-lipsync/pilot-compare-grid-2x2.mp4. Note the face is always full-bleed (hair top sits above the X/IG 4:5 crop, eyes and mouth inside). HeyGen output is 25 fps: re-encode to 30 fps. Do not use HeyGen for lines that need a calm, closed-lip delivery without testing. Check the HeyGen wallet before and after every run; never top up without Bandar/Emily. Pilot clips still get the P12 end card (Tamara only while Tabby is not live) and the three P21 platform cuts.
P23 31 Oct 2026 9:22 AM, pt575u ("Bring back Tabby. All posts.") Every end card shows BOTH the Tabby and the Tamara official logos in their real colours, with the pay note «نقبل الدفع عبر تابي وتمارا». This is Bandar's own instruction and SUPERSEDES the "do not publish we accept Tabby until Tabby is live" caveat in P12, open flag 3, the Tamara-only end card of 9:03 AM (P22 turn) and the Tamara-only wording of options 3 and 4 (9:20 AM). Applies to ALL posts (videos and images). Slogan stays «خلّ ورا ظهرك «سند»» on one line, smaller than the logo; palette #2121F5 / #FAF8EA / white only, the two logos on white chips in real colours; no URL. Canonical end card = v3 (brand-endcard/sanad-endcard-v3.png|mp4, layers layers-v3/): Tabby TabbyBadge.svg (official media kit) + Tamara "Wordmark Gradient" (official assets), unmodified. Platform fit: scale the layer bbox (x103–979, y344–1365) into each safe rectangle (TikTok 0.90, Instagram 0.96, X 1.0). Options 3 and 4 regenerated with both logos (brand-endcard/options-20261031/). Dropping either logo; recolouring logos; a pay note that names only one provider. Note: STANDING-RULES.md had no Tabby clause; it is now rule 4. The 30 Oct Tamara-only rule in X-pack copy (social/x/…approval-pack*.md, "no Tabby/Tamara/payment methods") is a separate copy rule, not changed here: ask Bandar if text posts should name Tabby too.
P24 31 Oct 2026 10:09 AM, pt584u ("15 approved") New male brand face Abdulaziz (عبدالعزيز, male-01), named by Bandar 31 Oct 2026 (voice TBD), approved after 4 rounds / 16 options (rounds 1-3 rejected: "None. Regenerate", then "doesn't look Saudi enough"). Hejazi Saudi man ~28, warm medium-brown skin, full short sharp-lined beard, glossy gelled black hair, rimless glasses, white thobe, Jawaher studio style (option 15). Hero reels-library/characters/male-01/option-15-APPROVED.png; identity brand-face/identities/male-01/IDENTITY.md. Nano Banana 2.1 9:16, 1.5 cr per still, 24 cr total for 16 options; 4 new faces per round. No Soul ID yet (25 cr, only on Bandar's request); use image-reference generation from the hero. For lip-sync use P22 (HeyGen Avatar IV) with a frontal chest-up still, hands out of frame. Passport-style stills and generic Middle-Eastern faces (rejected twice); prompts without explicit Najdi/Hejazi features and sharp-lined Saudi beard; hands or angled body in lip-sync stills.
P25 31 Oct 2026 10:18 AM, pt591u ("No need to repurpose… show me a video for approval. And pause.") One video for approval only; NO automatic platform cuts. Deliver a single 9:16 master to Bandar and stop. TikTok and X versions are made ONLY when Bandar says "repurpose", which runs the repurpose skill (TikTok + X, safe zones per 3b, named <project>-<video>-<Platform>-v<N>-<YYYYMMDD>.mp4, uploaded to Drive). Supersedes P21 until he says otherwise. Send the master, wait for approval, wait for "repurpose". Building TikTok/Instagram/X cuts automatically after approval.
P37 8 Nov 2026 5:22 PM, GO v2 Voice-model choice is UNRESOLVED (eleven_v4 vs eleven_v4_turbo). The approved A–F GO v2 confirmation says this needs an explicit choice rather than a silent switch. Do not treat either model as the only allowed default. Tabby logos on end cards (P23) stay a creative rule and are not a live-checkout claim; Tabby checkout is not live. The Social Media bot stays inactive and does no content development; Claude is the content manager with the Content Engine and is not an approver. Do not claim interview-prep is built. O7 (any monthly application cap, including a 300 figure) is UNDECIDED. Studio and the paused teams stay OFF. Follow v2/CURRENT-CLAIMS.md and v2/V2-ROLLOUT-STATUS.md. Using older voice-only, Tabby-live, 300-cap, or Social-Media-as-producer lines as current defaults.
P38 8 Nov 2026 5:58 PM, v2 docs Faisal is dropped for new content. Current faces for new work are Jawaher, Hessa and Abdulaziz. Faisal’s historical files stay on disk and must not be used in new posts, reels or generations. Documentation change only; no assets deleted. Use Jawaher, Hessa or Abdulaziz for new work. Using Faisal in new content after this date.

3. Lip-sync QA recipe (run on every talking segment before showing Bandar)

  1. Cut the segment from the final reel (ffmpeg -ss … -t …).
  2. Per frame: MediaPipe FaceLandmarker (/workspace/tools/face_landmarker.task, venv /workspace/tools/venv-qa, scripts reels-library/reel-1/test/v3/v6/m.py, sm.py): lip gap = |L14.y − L13.y| / face height; width = |L291.x − L61.x| / face height; blendshapes jawOpen, mouthSmileLeft/Right.
  3. Speech energy: VO resampled to 16 kHz, log-RMS in a ±25 ms window at each frame time.
  4. Print one row per 0.2 s: time, voice-line time, opening, width, dB. Pass = mouth open (> 0.04) in roughly 80%+ of loud frames, and moving (std > 0.02 per 0.5 s window), closures only in gaps; no open mouth in silence before the first word; smile curve low until the last word ends.
  5. Build the 0.2 s face contact sheet (28 tiles for a 5.6 s tail) and look at it. Metrics are noisy, the sheet decides.
  6. Also confirm the caption starts within ~0.1 s of its first word and that frames outside the edit are SSIM 1.000 against the approved file. Note: Sync's output file has an audio offset of about +330 ms in the container; ignore it, the reel plays the HyperFrames vo.wav. Mouth opens ~100 ms before the audio peak, which is normal.

3b. Platform safe zones, 1080×1920 (checked 30 Oct 2026; apply to every video AND image, P21)

What went wrong on reel/hero in Bandar's TikTok screen recording (30 Oct 2:05 PM, iPhone 1290×2796, business-account profile opened from a search): TikTok scales the 9:16 frame to the full screen height, trims about 50 px per side, draws a "Find related content" search bar over the top ~280 px, a location chip + username + «تعرف على سند» line from y≈1600 down, a promo banner at the very bottom, and the like/comment/save/share column from x≈910 down. Our burned-in captions at y≈1360–1470, the top-left logo / pill at y≈270 and anything near the right edge were covered or cut. | | Top clear | Bottom clear | Left / right | Right button column | Also | |---|---|---|---|---|---| | TikTok (design rectangle used: x 120–910, y 290–1440) | ≥ 280 px (search bar; official ad template 240) | ≥ 480 px (measured in-app: username/caption start at y≈1600, so 320; the CTA line and ads/boost banner add more; official ad template 660) | ≥ 120 px left; ≥ 170 px right | x ≥ 910 from y ≈ 840 down (about 140–170 px wide; official template says 300) | Boosted/Spark posts add a CTA button: treat the official 240 / 660 / 120 / 120 template as the strict zone if a video is ever promoted. | | Instagram Reels (design rectangle used: x 100–940, y 280–1500) | ≥ 270 px (Meta official: 269; status bar + Reels header) | ≥ 420 px (measured ~320–420; Meta official ad template 672) | ≥ 65 px official; we keep 100 | x ≥ ~940 from y ≈ 1150 down | Grid and feed crops: the profile grid shows the centre 3:4 (1080×1440, y 240–1680) and a Reel shared to the feed can show the centre 4:5 (1080×1350, y 285–1635): keep hook text, face and cover title inside y 285–1635. (Sources disagree on the exact grid ratio, so use the stricter 4:5 for anything essential.) | | X (design rectangle used: x 90–990, y 290–1470) | ≥ 80 px (community measurement), but keep text and the face below y 285 so they survive the 4:5 timeline crop | ≥ 450 px (community measurement; player controls along the bottom) | ≥ 80 px each side; we keep 90 | no side button column in the vertical viewer | Timeline crop: keep essentials inside the centre 4:5 (1080×1350, y 285–1635) as Bandar specified. X publishes no safe zone for organic video: the 80/450/80/80 figures are third-party measurements (superguides.pro, checked 23 Oct 2026, low confidence). X's own specs list 1080×1920 (min 720×1280), H.264 + AAC LC, 5–10 Mbps target, up to 2:20, but say nothing about how a standalone 9:16 video is cropped in the timeline; they only say Amplify Sponsorships crop tall video to 1:1 and an old page says 1:1 may crop to 16:9 on mobile. If X ever crops reel 1 to 1:1 (y 420–1500) the top of the face would be cut: use the 4:5 rule unless Bandar sees otherwise on his own timeline. | Rules: (1) keep every text card, caption pill, bio pill, logo and the end card (logo, slogan, Tabby/Tamara) inside the design rectangle; full-bleed backgrounds and the face video may bleed outside it. (2) Captions sit at y≈1320 (TikTok) / 1380 (Instagram) / 1350 (X), i.e. just above the bottom limit. (3) Never use black bars; if a re-render is not possible, scale the whole frame down onto #FAF8EA or #2121F5. (4) Re-check numbers against current sources each quarter; there is no official organic spec, only ad templates (TikTok In-Feed 240/660/120/120; Meta Reels 269/672/65/65) plus in-app measurements. Sources read 30 Oct 2026: quelta.ai/tools/safe-zones/tiktok, solidlabs.com/social-safe-zones, adchecklab.com/specs, shotcraft.video/reference/platform-safe-zones, creobee.com Instagram reel safe-zone page, cadenus.io Instagram safe zone, fooh.com Instagram grid.

4. Techniques and failures to avoid (technical)

5. Credits (observed, Starter plan, shared pool)

Unit costs: nano_banana_2_1 2k still ≈ 2 cr; Kling 3.0 std 720p ≈ 1.25–1.75 cr/s (6 s = 7.5; a 3-clip 15 s batch = 26.25); Seedance 2.0 Fast ≈ 2.5 cr/s; Wan 3.0 720p ≈ 1.75 cr/s; Sync Lipsync 3 ≈ 3.6 cr/s (5.5 s = 19.8; a 6 s run = 21.75). Soul ID training 25 cr each; reference-sheet images 1 cr each. Voice-over b-roll, UI cards, captions, end cards, save/share card: 0 Higgsfield credits. Reel 1 total ≈ 422.7 cr (model test 232.8, v2 58.5, v3 55.05, v4 21.75, v5 0, v6 54.6, v7 0). Identity library (one-off) ≈ 124.1. Balance after v6: 761.04. Per-reel planning figures for the other 14 reels are in PRODUCTION-PLAN.md section 1; update them with the real reel-1 figure (about 423 vs planned 73 with re-rolls: talking-head refinement is the expensive part, budget 2–3 refinement rounds). Reel 13 lip-sync pilot (31 Oct 2026, one 10 s line, four engines): A Kling3 + Sync3 = 53.5 Higgsfield credits; B OmniHuman 1.5 (fal) = $1.62; C Kling AI Avatar v2 Pro (fal) = $1.17; D HeyGen Avatar IV = $0.40 (chosen, P22). HeyGen wallet $5.00 → $4.60; fal balance not readable by the API key (B and C costs computed from billed duration × list price; another process may have run duplicate fal jobs on 31 Oct 08:39–08:46, so check the fal dashboard). Higgsfield balance after A: 462.39.

6. Assets index

7. CHANGELOG, reel 1 «سيرتك ترفعها مرة وحدة بس» (Jawaher, 36–37 s)

Version Date What changed Why Credits
Model test (Kling / Seedance / Wan) 30 Oct Same stills, prompts and VO on 3 models; same HyperFrames cards. Pick the closest model (P8). Kling won. 232.8
Kling v1 30 Oct First assembled version (Kling 3.0 std, Sync on S1 + S6, S3 over-the-shoulder hands). Baseline. in the 232.8
Kling v2 30 Oct New still-body Kling bases for S1 and S6 (silent, neutral mouth, negative prompts) + Sync on S1 (3 s) and S6 (9 s); S3 hands reused from v1. Mouth/body artefacts of v1. 58.5
Kling v3 (approved first half) 30 Oct Two expressive S6 bases with one timed presenting gesture; Sync on the chosen one (trimmed L6). Saved as sanad-reel1-kling-v3-APPROVED.mp4. Hand timing and expression (P1, P5). Bandar: "first half is perfect". 55.05
v4-candidate (+ v2 logos) 30 Oct 12:12–12:37 Attempted tail re-sync from 3.0 s (not used: timing identical to v3), added the signature end card, then real Tabby/Tamara logos on white chips. Bandar: second half after «أو» mismatched (P2); end card rule (P12). 21.75
v5-candidate 30 Oct 12:40 Smile delayed by a mouth-patch hold + crossfade; real-logo end card. Smile started before «لك» finished (P6). Failed: froze the mouth (P4); Bandar: "same issue". 0
v6-candidate → v6 APPROVED 30 Oct 12:46 → locked 1:25 PM Regenerated everything from «أو» to the end: Kling 3.0 std 6 s from frame 76, neutral-expression prompt; Sync Lipsync 3 with the full untrimmed L6 audio. Verified every 0.2 s (mouth open in ~83% of loud frames, moves in every 0.5 s window). First 18 s pixel-identical, first 36 s audio bit-identical. Real articulation (P2, P3), consistency (P1). Known: talking-smile starts before «لك» ends (P6), hand gesture later than approved take. 54.6 (attempt 1 discarded: 27.3)
v7-candidate 30 Oct 1:31 PM Built on v6 APPROVED, only the save/share card changed: new layout (bio pill on top, blue panel with bookmark + paper-plane buttons, single line «احفظ الفيديو وشاركه مع شخص يحتاج وظيفة»), bio cue kept, template saved. Everything outside 26.7–32.7 s verified SSIM 1.000, audio md5 identical. Bandar 1:30 PM redesign (P11). Spoken L7 still has the older wording; end card still has the old slogan (locked file). 0
v7 TikTok + Instagram cuts → APPROVED 3:03 PM 30 Oct 2:30 PM Same v7 (audio bit-identical, no lip-sync/voice/timing change); overlays re-rendered per platform: caption pills moved from y≈1360 to 1320 (TikTok) / 1380 (Instagram), badges to y 290 / 280, every card group and the end card scaled (0.84–0.98) into the safe rectangle, end card keeps the OLD slogan to match the locked L8 voice. H.264 High crf 14 (~5 Mbps), AAC copied. Files sanad-reel1-TikTok-v7-20261030.mp4, sanad-reel1-Instagram-v7-20261030.mp4. P21: Bandar said the video did not fit the TikTok screen. 0
v7 X cut 30 Oct 3:20 PM Same v7, audio MD5-identical, overlays re-rendered for X: x 90–990, y 290–1470, captions at y 1350, essentials inside the centre 4:5 (y 285–1635). H.264 High crf 14, AAC copied. File sanad-reel1-X-v7-20261030.mp4. P21 widened to X, Instagram, TikTok (pt556u). 0

8. Open flags

  1. Slogan: confirm the inner «سند» guillemets on first use of end card v3; reel 1 locked file keeps the old slogan on card and in voice (L8). A re-cut would need a new L8 (ElevenLabs only).
  2. Reel 1 L7 voice says «لصديق يدوّر وظيفة»; on-screen card now says «لشخص يحتاج وظيفة». Re-record L7 (ElevenLabs, 0 Higgsfield credits) if Bandar wants them identical.
  3. Tabby is not live: no publish of "we accept Tabby" until it is. SUPERSEDED 31 Oct 9:22 AM (P23, pt575u): Bandar brought Tabby back on every end card.
  4. Progress-ring look (P16) not yet shown to Bandar; Hessa/Jawaher final voices still to be picked; reel 1's NoraS_F is a library voice.
  5. PRODUCTION-PLAN.md credit table predates reel 1 actuals (about 423 vs 73): update before planning the next reel.
  6. The TikTok screen recording Bandar sent (pt547u) shows the HERO video (dashboard/phone-mockup «قدّمت على وظايف…», uploaded to @sanadwa3e), not reel 1. The same margin problems apply; a TikTok/Instagram cut of the hero (sources under hero-videos/projects) is still to be made if he wants it.

Approval log update · 30 Oct 2026, 3:03 PM (Riyadh)

Approval log update · 31 Oct 2026, 9:22 AM (Riyadh)

Approval log update · 31 Oct 2026, 9:03 AM (Riyadh)

X (Twitter)

Approved by Bandar 30 Oct 2026 (3:07 PM Riyadh). Source pack: /workspace/sanad/social/x/20261030-x-approval-pack-FINAL.md (cleared by Copy Humanizer and Emily). Posting queue: /workspace/sanad/social/x/posting-queue.json.

Approved profile

Rules

  1. Account: free account only (no Premium, no ads, no spending). Actions are by hand or in a human-approved session; no auto-replies, DMs, follows or bots. Nothing public before Bandar approves the copy.
  2. Brand: #2121F5 / #FAF8EA / white only, locked wordmark, no neon. Avatar 400×400 inside the circle crop.
  3. Text posts: one idea per post, ≤ 280 characters (a URL costs 23), hook in the first line, a clear next step (repost, reply, vote), 0–2 hashtags (0 by default; #وظائف or #وظائف_السعودية only when it fits). One emoji at most as a pointer.
  4. Threads: 5–7 posts, a hook post that stands alone, one idea per post, CTA and slogan in the last post. Any example person is labelled fictional inside the thread. Default last post is the no-price 7/7.
  5. Polls: 4 options max, each ≤ 25 characters, 24 hours. Questions are about the audience's job-search experience, never a claim about سند. Publish real results in words afterwards, no invented numbers.
  6. Video: MP4/MOV, H.264 + AAC, 30 fps, 9:16 at 1080×1920 (default), 1:1 only if built for it. Free limits: 140 s, 512 MB. Aim 15–40 s, burned-in captions (autoplay is muted), hook in the first 2 s. Reel 1 file for X: reels-library/reel-1/sanad-reel1-X-v7-20261030.mp4.
  7. Safe margins (9:16 1080×1920): key text and faces inside x 65–1015, y 230–1632. Keep the bottom 15% (y > 1632) and the top ~12% clear of key text. No URL on screen. On X «احفظ» = bookmark, «أرسله» = repost; wording «الرابط في البروفايل» and «أعد نشره لصديق يدوّر وظيفة». Reel 1's voice says «البايو»: accepted by Bandar, no re-record.
  8. AI label: AI-generated faces carry the on-screen chip «مولّد بالذكاء الاصطناعي» the whole video and the line «شخصية مولّدة بالذكاء الاصطناعي. القرار النهائي عند جهة العمل.» in the post text. Characters are never shown as real clients or staff.
  9. Tone: Saudi Arabic, plain and warm, like a friend who knows the job market. Short sentences. No hype, no guarantees («سند ما يضمن وظيفة»), no invented results, stats or testimonials. Don't shame the job seeker.
  10. Claims: only what's built: CV upload, job matching, a tailored CV per job, apply only on jobs the user approves or auto-apply turned on for best-fit jobs, and the dashboard. Apply wording exactly: «سند ما يقدّم إلا على اللي توافق عليه، أو تفعّل التقديم التلقائي فيقدّم عنك على الوظائف الأنسب لك.» No government platforms or Nafath, and don't explain who applies (Bandar's rule, to Emily, 29 Oct). No Tabby, Tamara or any payment method until Bandar clears it and it is live.
  11. Price is held unless Bandar approves showing it: post 4 and the priced thread 7/7 stay out of the queue.
  12. Links: the website lives in the profile field and posts say «الرابط في البروفايل». No link replies by default; only if Bandar approves one, the URL goes in the first reply, never in post text or a video.
  13. Schedule (Riyadh, starting hypothesis, not data): Sun–Thu mornings 8:30, lunch 12:30, late evening 9:30 PM; Friday and Saturday lighter. One or two posts a day at most; reply by hand to genuine replies within the day. Review analytics after two weeks.
  14. Process: copy → Copy Humanizer → Emily → Bandar. Log every post with its date and link in the changelog.

Changelog

Date Change
30 Oct 2026 X pack approved by Bandar. Header regenerated with the group's visible-pixel bbox centred exactly on (750, 250) (v1 kept). Posting queue created.

Approval log update · 31 Oct 2026, 10:09 AM (Riyadh)

Approval log update · 31 Oct 2026, 10:18 AM

Approval log update · 31 Oct 2026 (Riyadh), reel 14 Hessa remake

Approval log update · 31 Oct 2026, ~10:55 AM (Riyadh), reel 14 v2

P26 — Shared voice for Hessa and Jawaher (Bandar, 31 Oct 2026, 11:00 AM Riyadh) — LOCKED

P27 — No on-screen AI label (Bandar, 31 Oct 2026, 11:04 AM Riyadh)

P28 — Slogan + sonic logo (Bandar, 31 Oct 2026, 11:04 AM Riyadh) — design pending approval

P29 — TikTok link-fix copy (Bandar via Emily, 31 Oct 2026, 11:14 AM Riyadh)

P30 — Sonic logo + no written URL (Bandar, 31 Oct 2026, 11:23 AM Riyadh) — APPROVED

P31 — Head-to-shoulders default framing (Bandar, 31 Oct 2026, pt605u)

P32 — Collision check is mandatory (Bandar, 31 Oct 2026, pt604u "Same overlapping issue")

Final end card + reel 1 v8 (31 Oct 2026, P23/P27/P28/P30)

P33 — Reel 1 v8 and all older Reel 1 cuts RETIRED (Bandar, 31 Oct 2026, pt613u) — Reel 1 rebuilt as v9

P34 — Reel 1 v9 RETIRED (Bandar, 31 Oct 2026, ~11:50 AM Riyadh) — Reel 1 rebuilt as v10

P35 — Reel 14 v5 APPROVED and LOCKED; Hessa's look locked (Bandar, 31 Oct 2026, 12:36 PM Riyadh)

P36 — No stills (Bandar, 3 Nov 2026)

Never use still images as shots. Every presenter shot is real moving footage filmed as one continuous setting, and inserts and cards play moving clips.

Reel A v2 built (PENDING Bandar) — 3 Nov 2026, 12:26 PM Riyadh (Abdulaziz, «سند عنده الحل»)

9. 8 Nov 2026 operating overlay

P37 added after GO v2 (staged deploy) and the content-manager correction. Older preference rows are not rewritten. Current claims and holds: v2/CURRENT-CLAIMS.md.