gooseworks-ai/goose-skills / skills/video/render-street-interview
render-street-interview
VerifiedWrite a street-interview ad from a complete inspected commercial interaction and current brand facts. Render product guessing, or prepare mic-only, prepared-sample or visible-task conversation script/prompt previews. Premise, actions and ad connection vary; capture rules stay fixed. Conversation media delivery is unverified.
tests/READINESS.md
View on GitHub# Human version
**Summary.** This is a dated readiness audit from 2026-09-29, not today's runbook. The 2026-10-01 face verdict in `REFERENCE.md` supersedes its unruled-face blocker, but does not prove a new take's quality. Current execution uses `SKILL.md`, the selected recipe and [bundled model notes](references/model-behaviors.md).
The private paths, missing measurement guide, footage and proposed experiments below are optional historical authoring evidence. No Studio checkout or avatar still is required. Default zero-strength finishing needs no colour reference. Old cost figures and classifier observations are not current pricing or policy guarantees.
---
# Agent version
# create-street-interview-video — readiness
**Current summary (2026-10-03).** Product-guess retains the existing object renderer and
approved prompt baseline. Conversation supports **free script/config/prompt previews only**
for mic-only, prepared-sample and visible-task interactions. No conversation media endpoint,
price, performed timing, sampling action or camera/audio result is validated. Every subtype
still refuses `single_gen.py --yes` and product/scene reference bindings.
**Prompt version 3 (2026-10-06): draft.** v3 bounds the background falloff so the street stays
in focus, never blurred and never bokeh. Versions 1 and 2 are byte-identical to before. v3 has
passed only free checks (lint, snapshots). It stays draft until one 720p take, separately
approved with its quoted price, measures inside the real-footage detail band (4.76-9.60) with
no bokeh.
## Current conversation preview support
- Configs may describe `interaction.type`, `visible_setup`, `participant_reason` and optional
props. Missing interaction data remains mic-only. Product-sample uses a prepared plain cup;
concept-challenge uses described non-UI props; mic-only can explain a service without a
device. Phone, screen and UI demonstrations remain unsupported.
- Ordered shots can include explicit visual/action descriptions and silent action/reaction.
A participant answer or silent action can open the edit. The first actual interviewer
question exactly matches `cfg.question` and is spoken once. Both people still speak.
- The native preview checks retain 3–8 shots, a 6–15s duration and the provisional ceiling of
2.5 spoken words/s. Authors must leave breathing room for actions; these checks cannot
verify performed timing or quality.
- Writing starts with a complete inspected commercial street interaction and a coherent
situation brief. Full visual and spoken coverage, setup, participation, hook, product
role and payoff are required. Seed snippets are leads; radio/editorial material is not
a fallback. Unseen recruitment stays unknown. See
[street-script-writing](references/street-script-writing.md).
- Human review is on the situation brief, words, action timeline, earned brand connection,
ending-card specification and prompt preview. A clean dry run verifies construction,
not media delivery. A rendered pilot and delivery review are still required before
conversation can become a paid production path.
## Historical product-guess readiness (2026-09-29)
The report below is preserved as history, not the current production verdict. The later
[REFERENCE.md status](REFERENCE.md#status) records that the user accepted the single-take
faces on 2026-10-01 and that a product-guess episode shipped. That does not validate any
conversation subtype. Product-guess prompt instructions remain unchanged.
State as of 2026-09-29. Written from a pass over the skill with **no paid generation of any
kind**: no still, no clip, no voice. Everything measured below came out of the take already paid
for (`gw-single-seed4802`) and the local reference footage.
---
## What now runs unattended
| Script | State | Verified |
|---|---|---|
| `paths.py` | new | repo root found by marker, run folder via `--run` / `$STREET_INTERVIEW_RUN` / default |
| `single_gen.py` | dry run clean, no network | prints model, seed, price, destination and the full prompt; `fal_helpers` import is now lazy so a dry run needs no key |
| `variants_gen.py` | dry run clean, no network | 6 variants, seeds 4810–4815, all paid (`--fast` is the cheaper endpoint) |
| `build_looks.py` | `--dry-run` clean, and a **real free render verified end to end** | dry run renders every caption layer and asserts the safe zone with no ffmpeg; the real run produced `street-gw-bare.mp4` 14.30s and its caption-free control |
| `phone_look_video.py` | runs; now fails loudly instead of silently | zero-strength finishing needs no reference; positive-strength matching requires a real reference |
| `check-cut.py` | **new gate, 9 checks, falsified both ways** | full PASS on the `bare` look with every check live, including Whisper |
Run order, all of it free except step 2:
```bash
cd skills/molecules/create-street-interview-video/scripts
python single_gen.py # dry run: prompt + price, nothing sent
python single_gen.py --yes # PAID; runs only on an approved order
python build_looks.py --dry-run # free, no ffmpeg
python build_looks.py # free, all five looks + controls
python check-cut.py --look subway --falsify # prove the gate can fail
python check-cut.py --look subway # the ship gate
```
### What was broken before this pass
Three of the four scripts crashed on their own dry run from the skill directory. All of it was
path resolution: the scripts were lifted out of `projects/street-interview/working/`, where
`Path(__file__).parents[2]` happened to be the repo root and `./takes/` happened to hold the
take. Inside `skills/molecules/create-street-interview-video/scripts/` `parents[2]` is `skills/`
and `./takes/` does not exist.
- `single_gen.py` — `ModuleNotFoundError: fal_helpers` **on the dry run**, because the import was
at module level. A dry run that needs an API key is not a dry run.
- `build_looks.py` — looked for the take at `scripts/takes/gw-single-seed4802.mp4` and wrote its
caption-free controls into `scripts/graded/`, i.e. into the skill directory, breaking the
run-folder convention outright.
- `check_cut.py` — same, plus it defaulted to `HERE.parents[0]/output/looks/`, and its
`ffprobe` failure surfaced as `KeyError: 'streams'` rather than a message.
- `phone_look_video.py` — its colour reference defaulted to an Arcads reference clip in a
refs/tools folder resolved inside the skill, where nothing exists.
- `variants_gen.py` — the only one that dry-ran, and only because it defers the fal import. Its
`CAN` path was still wrong, so `--yes` would have failed at upload after printing a price.
Also fixed: the gate was named `check_cut.py`, against the `check-*.py` convention every other
molecule here uses (`check-episode.py`, `check-timing.py`). It is `check-cut.py` now.
### What is still missing, and is not a code problem
1. **`projects/` is gitignored, so the entire knowledge base is invisible on a fresh checkout.**
The reference footage, the 40KB run log, the approved take, the five finished looks and the
colour target all live under `projects/street-interview/`. `SKILL.md` says "Reference build:
`projects/street-interview/`. Copy it, do not blank-page a new one" — which nobody cloning
this repo can do. Mitigated here by copying the measurements into
`references/REFERENCES.md` and the rejections into `TAKES.md`; **not** mitigated for the
footage, which has to be re-downloaded.
2. **`SKILL.md` pointed at `AVATAR_GENERATION.md` with no path.** It is
`one-shot-videos/AVATAR_GENERATION.md`. Fixed.
3. **`skills/atoms/_shared/MODEL_BEHAVIORS.md` carries a stale entry** — "Seedance 2.0 · the
likeness gate makes it unusable for photoreal street strangers (2026-09-28) … route it to H3
or Veo from the start". Seed 4501 disproved it: the gate inspects **uploaded images**, not
generated faces, so `text-to-video` and `reference-to-video` with a *product* reference both
work. I have deliberately **not** edited that file: it has uncommitted changes from another
session running in this same checkout (`SYNC_LEDGER.json` and `SYNC.md` disappeared from the
working tree while this pass was in progress). One correcting entry is owed there and should
be made by whoever owns that file's current edits.
---
## What is unproven
### 1. The blocker: nobody has ruled on whether the approved take's faces pass
This is the one that matters and it is not a code gap. Four rounds of generated street strangers
were rejected for reading as AI (fal stills → Veo/H3; two still re-rolls; H3 with a real
reference video). The standing rule is that **a passing example must exist before more spend.**
But the four rejections were all on the *pre-single-generation* routes. `gw-single-seed4802` is
recorded as "the approved base" and was approved on structure — one location, one room tone, a
mic in every shot. **No verdict has ever been given on its faces specifically.** Until that
verdict exists, every priced variant in `variants_gen.py` is a bet on an untested assumption,
and the recipe is not production-ready no matter how clean the scripts are.
### 2. The pace gap, which is measured and unclosed
| | median shot | shots |
|---|---|---|
| Salary Transparent Street | 1.60s | 37 in 74.4s |
| Chris Klemens | 1.62s | 12 in 22.4s |
| SubwayTakes | 1.54s | 64 in 122.2s |
| **ours (seed 4802)** | **2.61s** | **4 in 12.1s**, final shot 5.18s |
Three references across two frame rates and 220s of material agree inside 150ms. We are ~60%
slower, with a final shot longer than any shot in Klemens. This is free to change — it is how
many internal shots one call is asked for — and it has never been tried.
### 3. Nobody has heard any of it
Every audio judgement on this format is measured, never listened to. The loudness, the per-shot
noise floor and the Whisper transcript all pass; that is not the same as it sounding right. Say
so when handing a cut over.
### 4. Untested endpoint
`bytedance/seedance-2.0/image-to-video` is confirmed live on fal and is what the best-in-field
example uses. Its likeness gate may behave differently from `reference-to-video`. Never tried.
### 5. One brand only
Everything is built against Grimwater, whose joke (a can that looks like beer and is water)
is unusually well suited to a vox pop. Greenshot is the outstanding second brand and would show
whether the recipe generalises or whether it is one good joke.
### 6. What `check-cut.py` cannot see
It cannot tell whether a face reads as AI — no automated check caught any of the four rejections
— and it cannot tell whether a viewer can follow the video. Check D gets as close as a payload
can (the product must appear in shot 1 and the final shot, after a cut was once rejected for
showing the can only at 19s), and a human still has to watch the whole thing (`/watch`).
---
## The cheapest next experiment
### Do this first: it is free
**Ask for a realism verdict on the approved take, side by side with the real references.**
Build a compare video: `street-gw-bare.mp4` (no captions — the honest control, graphics cannot
carry it) cut against `refs/sts.webm`, `refs/klemens.webm` and an Arcads clip (real
licensed actors) at matched size, and ask one question: *which of these read as AI?*
- **Cost: free.** ffmpeg only; every input already exists on disk.
- **Time: minutes.**
- **What it settles:** whether the single-generation route cleared the bar that four earlier
routes failed, which is the exact precondition the standing rule demands before more spend.
- **Why it is the right next move:** Six paid variants are written. Sending any of
them before this verdict is spending on an assumption that has been wrong four times. If the
answer is "still AI", the honest conclusion is the one already in `SKILL.md` — Arcads-grade
realism is ~1,000 licensed real actors on a paid tier, which is a licence, not
a prompt — and the correct next step is a purchase decision, not a re-roll.
### If and only if that verdict is "this passes"
**One `--fast` call to close the pace gap.** Same payload as seed 4802, new seed, asking for 4
internal shots inside a 6s take instead of 12s, which targets the references' 1.5–1.7s median.
| | |
|---|---|
| endpoint | `bytedance/seedance-2.0/fast/reference-to-video` |
| duration | 6s (per `CLAUDE.md`: generate ~6s, not 8s — defects live in the unused tail) |
| resolution | **720p** (per `MODEL_BEHAVIORS.md`: the classifier sweeps harder at 1080p; preview at 720p first) |
| seed | pinned and recorded in `TAKES.md`, as every seed here is |
| **cost** | paid; the cheapest call here (6s on the fast endpoint) |
| fallback if only 4s/8s durations are accepted | a 4s or 8s take on the same endpoint |
| for comparison | a fraction of one full-length 12s take, and far less than the six written variants |
It is a real test rather than a re-roll because it changes one measured variable — shot count per
second — against a number taken from three references that agree. If 4 shots in 6s holds one
location and one room tone, the pace gap closes with one short call and the recipe is production-ready. If
6s cannot hold four shots, that is worth knowing for one short call and is recorded, not repeated.
**Each paid step runs only on an approved order.** The lock order applies: the free
verdict is the cheap artifact, and it comes before the short paid call, which comes before
the six-variant sweep.
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 |