fix-my-look
Edit the source's first usable frame with gpt-image-2 from the user's prompt,
then propagate that look across the clip with kling reference-video while
locking the original face, motion and audio via the original video + audio as
references. All prep happens in one normalize_video
call for short clips, or one normalize call per segment for longer clips. The
output ratio uses the normalized clip's closest supported output ratio; this
skill does NOT reframe the source video.
Inputs
<source>— path or URL to a video file with audio<change_prompt>— what to change (e.g. "make it night with neon lights", "change my shirt to a leather jacket", "put me on a beach in Hawaii")
Empty-args menu
- "What's the source video path?"
- "What do you want to change? (e.g. 'put me on a beach', 'make it night')"
Workflow
Working dir: ~/Downloads/fix-my-look/<run-id>/.
Step 0 — Cost, timer and task IDs
Every tool named below is a Pika MCP tool, written by its bare name. Call it under whatever prefix your session exposes for the Pika MCP.
Start a timer when the source and change prompt are known. Before paid
generation, call estimate_cost for the planned
generate_image_edit,
generate_reference_video, any multi-segment
edit_concat, and any optional audio/lipsync repair
call. If cost is not surfaced by the host, say
Cost not surfaced by this harness in the final report instead of guessing.
When any tool returns a
task_id, copy the exact value into the run notes and reuse it verbatim; do not
hand-type long JWT-style task IDs.
Step 1 — Prepare the clip
Local file? upload_asset it first; an HTTPS media URL
passes directly. Decide the source windows before normalizing: use one 14.8s
window for sources <=15s, and split longer sources into ordered 14.8s windows.
Call
normalize_video(video_url=<source>, start_s=<offset>, max_duration_s=14.8, extract_audio=true, extract_face_frame=true)
once per window. Use the first window's face_frame_url for the edited still;
use each window's video_url as that segment's motion/identity reference. For
multi-window clips, also call
extract_audio_from_video(video_url=<source>) so the
final merged output can be restored to one continuous source audio track.
Wire the result into the rest: face_frame_url is the Step 2 edit target;
each normalized video_url is Kling's reference for that segment in Step 4;
set aspect_ratio = result.aspect_ratio ?? result.closest_aspect_ratio for
each normalize result, then carry that local aspect_ratio through the image
and video calls. If neither field is present, stop and report that normalization
output is missing an aspect label. Compute
duration = max(4, min(15, round(duration_s))) per segment, and use
resolution="720p"
unless the user asked for high res. If face_found is false, no clear face was
found and face_frame_url fell back to the t=0 frame — proceed but warn
identity may drift, or re-run with a start_s at a section where the subject
faces camera.
Reference-video providers can reject oversized reference assets. If the
normalize result or the downstream provider error shows a normalized video is
over the provider limit, retry normalize_video once
with crf=28 and the same start_s, max_duration_s, extract_audio, and
extract_face_frame values. If the reference is still too large, stop before
another paid video attempt and report that
normalize_video needs a worker-side 1080-edge /
reference-size cap. Do not patch this with local shell media commands.
Step 2 — Edit the frame with gpt-image-2 (the "change" stage)
generate_image_edit with provider="gpt-image-2",
aspect_ratio=<aspect_ratio>, resolution="2K",
images=[<face_frame_url>], quality="high", prompt:
"Modify the reference photograph as follows:
<change_prompt>. Keep the person's face, identity, hair, body and pose EXACTLY as in the reference. CRITICAL: preserve every object the subject is holding or touching — phones, products, drinks, bags, props, jewelry — in the exact same hand, position, orientation and scale; never remove, replace or restyle them. Change only the requested scene, background, clothing, lighting or environment, not who the person is."
Keep the "preserve held objects" clause verbatim on every re-render — without it gpt-image-2 silently drops products/phones the subject is holding.
If gpt-image-2 returns a content-policy false positive for fashion, glam, or beauty prompts, retry once with the same intent but a modest / editorial wording such as "polished event styling, opaque clothing, natural pose, non-sexual fashion portrait". For makeup prompts, explicitly preserve the original eye shape, eyelids, iris color and gaze; heavy eyeliner/eye shadow is a high-risk identity-drift source.
Step 3 — Show the edited frame and wait for approval
Surface the edited frame and STOP. Ask "Approve for video generation, or tweak and re-render?" Do NOT call video generation until approved. For tweaks, re-run Step 2 (locked clauses verbatim) and loop.
Step 4 — Propagate via Kling reference-video
For each normalized segment, call generate_reference_video
with provider="kling", reference_videos=[<segment video_url>],
reference_images=[<edited_frame_url>], aspect_ratio=<aspect_ratio>,
duration=<segment duration>, sound=false, video_keep_sounds=[true],
prompt:
"Apply the change shown in <<<image_1>>> to <<<video_1>>>. Keep the person in <<<video_1>>> with the EXACT same face, identity, expressions, motion and timing; preserve the original video's kept sound track. The new scene/background/clothing/lighting should match <<<image_1>>>. CRITICAL: preserve every object the subject is holding or touching in <<<video_1>>> — phones, products, drinks, bags, props — in the same hand and orientation every frame. Keep mouth motion active through the final frame when the person is speaking. Do not alter the person's identity."
Append any extra creative direction (e.g. "very cinematic, soft golden light") after the locked text — never replace it.
Do not pass sound=true to Kling with a video input. Kling rejects that
combination with error:1201 sound on is not supported with video input; use
sound=false plus video_keep_sounds=[true] to keep the source video's audio.
If the source was split into multiple windows, call
edit_concat(video_urls=[<segment outputs in order>]).
After concat, run
edit_audio_replace(video_url=<concat_url>, audio_url=<full_source_audio_url>, duration_policy="video")
when the merged output audio is missing, drifted, or discontinuous.
Only try Seedance if the user explicitly asks for it, or if Kling fails and a second provider attempt is useful. Use the same segmenting rule and record the provider error plainly if Seedance rejects the input or drops speech/action.
Async handling: if any call returns a {task_id, status} envelope, poll
task_status({task_id}) in a tight loop until terminal.
Step 5 — Audio, duration and identity QA
Before reporting success, verify the generated video against the source:
- Duration must not be meaningfully cut off. If output duration differs from the intended source window or merged source duration by more than 0.5s, mark the run as failed / needs follow-up.
- If the source has speech, audio must be present through the tail and mouth
movement must not freeze before the spoken content ends. If words are missing,
garbled, silent, or visibly out of sync, do not call the run
PASS. - The approved frame corrections must persist into the video. If the provider reintroduces a removed artifact such as eyeglass glare, mark it as a propagation caveat or re-render from a stronger approved frame.
- Compare identity at start, middle, segment boundaries, and end. If Kling preserved motion but changed the face, call that out as a provider limitation instead of a pass.
If the video is visually acceptable but speech audio is missing, incomplete, or drifted, offer one paid repair pass:
edit_audio_replace(video_url=<generated_video_url>, audio_url=<full_source_audio_url or segment_audio_url>, duration_policy="video")edit_lipsync(video_url=<audio_restored_url>, audio_url=<full_source_audio_url or segment_audio_url>, variant="v2-pro")
If the model froze the mouth near the end, do not keep escalating to sync-3
automatically; lip-sync cannot reliably recover a face track with no mouth motion.
Offer trim / regenerate instead.
Step 6 — Download + return
Download the result to ~/Downloads/fix-my-look/<run-id>/result.mp4 and return
that path plus the final report fields: source, edited frame URL, final video
URL, provider, job/task IDs, cost estimate or not surfaced, elapsed
time, QA notes, and follow-up issue.
