gooseworks-ai/goose-skills / skills/video/ugc-fixloop
ugc-fixloop
Verifiedthe UGC fix-loop toolkit — surgically re-render a bad window/beat of a single-take UGC master (stitch_replacement.py, pure FFmpeg) and GPT cross-model review a Seedance prompt before render (vet_seedance_prompt.py, routed through the openai-proxy). Fetch it into a one-shot UGC recipe so both scripts resolve on any machine and the vet call charged to your GooseWorks credits.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 3.4 KB |
| scripts/stitch_replacement.py | script | 6.4 KB |
| scripts/vet_seedance_prompt.py | script | 5.9 KB |
| skill.meta.json | config | 221 B |
| tests/smoke-test.md | markdown | 1.7 KB |
scripts/vet_seedance_prompt.py
View on GitHub#!/usr/bin/env python3
"""Cross-model review of a Seedance 2.0 reference-to-video prompt by GPT — a
deliberately NON-Claude second opinion before you spend the render.
Portable: routes through the GooseWorks **openai-proxy** (charged to your GooseWorks credits, real
key never touches this machine) — never a direct api.openai.com call. Creds: the cloud
sandbox's GW_MEDIA_PROXY_TOKEN + GW_API_BASE env when set, else
~/.gooseworks/credentials.json (the CLI writes it). Takes the prompt to review
as an ARGUMENT (nothing baked in).
vet_seedance_prompt.py --prompt-file working/seedance-prompt.txt \
[--brief "one-line intent"] [--refs "@Image1=avatar; @Image2=product; @Image3=env"] \
[--words 28] [--model gpt-5.5] [--out working/seedance-review.md]
Exit 0 + prints/saves the review. Exit 3 if the proxy/creds are unavailable (so the
recipe can fall back to an inline self-review — the review is advisory, not a gate).
"""
import argparse
import json
import os
import pathlib
import sys
import urllib.request
import urllib.error
from urllib.parse import urlencode
SYSTEM_PROMPT = """\
You are a senior AI-video prompt engineer who has shipped hundreds of short-form UGC
ads on ByteDance Seedance 2.0, Kling, and Veo. You know Seedance 2.0 reference-to-video
intimately: how it binds reference images addressed as @Image1/@Image2/@Image3, native
lip-synced dialogue (generate_audio=true), where it drifts (identity, product geometry,
small text, multi-cut consistency over a single ~15s render), and how spoken-word budget
maps to clip duration. You have strong opinions and you don't hedge — when you spot a
problem you name it and give the specific rewrite, not vague encouragement."""
USER_TEMPLATE = """\
Review this Seedance 2.0 reference-to-video prompt for a single vertical (9:16) UGC ad
before we render it. {brief}{refs}Native lip-sync means a tight spoken-word budget —
our rule of thumb is ~{words} words for a ~15s clip; push back if you disagree.
# The Seedance prompt to review
---
{prompt}
---
# What I want (be concrete; don't soft-pedal — if it will produce mush, say why)
1. **Verdict** — ship-as-is / ship-with-tweaks / needs-rewrite, one sentence.
2. **Specific line edits** — quote the phrase, give the replacement, why in <=15 words. Focus on
what changes Seedance's output: @Image ref binding, camera-switch language, how hard cuts are
signalled, consistency anchors, wording that invites drift. 0-8 edits.
3. **Word/duration budget** — will every spoken line land at ~15s with the cuts, or cut a line? which?
4. **Consistency risk** — rank the top 3 things most likely to drift across cuts (face, colors,
product geometry, small text, camera switch) + the single best prompt-side mitigation for each.
5. **Reusability note** — 1-2 rules to bake into the general prompt recipe so the next brief inherits them."""
def _cfg():
# Cloud sandbox (coworker chat): a per-session proxy token that already binds the
# billing agent — no agent_id needed. Same resolution as media_proxy.py.
env_tok = os.environ.get("GW_MEDIA_PROXY_TOKEN")
if env_tok:
base = os.environ.get("GW_API_BASE")
if not base:
fal = (os.environ.get("GW_FAL_PROXY_URL") or "").rstrip("/")
i = fal.find("/api/internal/")
base = fal[:i] if i > 0 else None
if base:
return base.rstrip("/"), env_tok, None
p = pathlib.Path(os.path.expanduser("~/.gooseworks/credentials.json"))
if not p.exists():
sys.exit(3) # no creds → recipe falls back to inline self-review
c = json.loads(p.read_text())
return c["api_base"].rstrip("/"), c["api_key"], c.get("agent_id")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--prompt-file", help="path to the Seedance prompt to review")
ap.add_argument("--prompt", help="the Seedance prompt inline (alternative to --prompt-file)")
ap.add_argument("--brief", default="")
ap.add_argument("--refs", default="")
ap.add_argument("--words", type=int, default=28)
ap.add_argument("--model", default="gpt-5.5")
ap.add_argument("--out", default="working/seedance-review.md")
a = ap.parse_args()
prompt = a.prompt or (pathlib.Path(a.prompt_file).read_text() if a.prompt_file else "")
if not prompt.strip():
sys.exit("give the prompt via --prompt-file or --prompt")
api_base, tok, agent = _cfg()
user = USER_TEMPLATE.format(
prompt=prompt.strip(), words=a.words,
brief=(a.brief.strip() + " ") if a.brief else "",
refs=("Reference slots: " + a.refs.strip() + ". ") if a.refs else "",
)
body = {"model": a.model, "messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user},
]}
pid = os.environ.get("GW_PROJECT_ID")
q = urlencode({"token": tok, **({"agent_id": agent} if agent else {}),
**({"project_id": pid} if pid else {})})
url = f"{api_base}/api/internal/openai-proxy/v1/chat/completions?{q}"
req = urllib.request.Request(url, data=json.dumps(body).encode(),
headers={"Content-Type": "application/json"}, method="POST")
try:
with urllib.request.urlopen(req, timeout=180) as resp:
out = json.loads(resp.read().decode())
except urllib.error.HTTPError as e:
print(f"HTTP {e.code} {e.reason}\n{e.read().decode(errors='replace')}", file=sys.stderr)
sys.exit(3) # proxy error → advisory step, let the recipe self-review inline
except Exception as e:
print(f"openai-proxy unreachable: {e}", file=sys.stderr)
sys.exit(3)
review = out["choices"][0]["message"]["content"]
outp = pathlib.Path(a.out)
outp.parent.mkdir(parents=True, exist_ok=True)
outp.write_text(review)
print(f"[vet] model={out.get('model')} usage={out.get('usage', {})} saved={outp}\n")
print(review)
if __name__ == "__main__":
main()
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |