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.
references/reference-pipeline.md
View on GitHub# Street interview: pipeline, looks, cost and quality checks
**Summary.** The stage-by-stage pipeline (which stages are paid and which are free), the five finishing looks, what costs money, and the quality checks every cut must pass.
Part of the street-interview development record, split out of [`REFERENCE.md`](../REFERENCE.md). It is history, not today's runbook: `SKILL.md` and the selected recipe win where they differ. Brands, seeds and props are those experiments' (fictional brand names), never another brand's defaults. File names in backticks (`TAKES.md`, `selftest.py`, `brands/...`) are relative to the skill folder.
---
## Pipeline
```
brand-research real facts and real assets [required first]
write 4 beats question, two wrong, one right [fits 12-15s or it does not ship]
brands/<slug>.json the creative, as data; copy the demo [free; brandkit.py validates it]
selftest.py the format survived the split [free; no network, no ffmpeg]
single_gen.py ONE call; dry run prints price + prompt [PAID, 12s/720p]
or variants_gen.py six alternative grammars [REFUSES: all six fail the lint]
fit_grade.py SOLVE the grade for THIS take [free; --write-brand stores it]
build_looks.py grade + brand layer, 5 treatments [free; --dry-run needs no ffmpeg]
recut.py drop the dead air; pace [free; writes a .plan.json map]
check-cut.py the ship gate; --falsify first [free]
falsify-all.py plant a fault for EVERY check [free; run it after touching one]
measure-pace.py shot lengths off the render [free; the gate does not see pace]
bench.py every take AND every real reference, [free; no network]
same axes, same code, one table
```
`build.py` chains generate -> grade -> recut -> gate in one command, and its output is a graded
re-cut TAKE, not a finished ad: nothing in that chain draws a caption, a title or an end card.
The brand layer is `build_looks.py`, and it runs from the UN-recut take. Two pipelines, and only
one of them finishes a video. See Critical knowledge 20-22.
`bench.py` is the benchmark, and it exists because every realism number this format has quoted
was measured by a different ad-hoc command on a different subset of files. It measures our takes
and the real references on one set of axes -- shot-length distribution, cuts per second, detail at
a FIXED pixel width, black point, face-region saturation, per-shot ambience floor, LUFS, true peak
-- and prints the band each row is in or out of, plus the band **re-derived from the reference rows
in that same run**, so an inherited band that does not reproduce is visible instead of applied.
Run it before believing any claim in this file.
Two layers, and neither holds the other's job:
| | |
|---|---|
| `scripts/format_spec.py` | the prompt scaffold, the shot grammar, and the clause list the gate lints for. Format knowledge. No brand appears in it. |
| `brands/<slug>.json` | product and its reference photo, location, question, cast, dialogue, props, seed, caption copy, series header, logo, end card, measured cuts, measured ambience window. |
`check-cut.py` imports its clause list from `format_spec` rather than keeping a copy, so a clause
cannot be deleted from the prompt without the gate noticing, and `single_gen.py`'s dry run runs
the identical lint before any spend.
Every script takes `--brand <slug>` (default `grimwater`, or `$STREET_INTERVIEW_BRAND`).
Every script resolves the repo root by marker and the run folder by `--run`
(default `projects/street-interview/`, or `$STREET_INTERVIEW_RUN`). See `scripts/paths.py`:
none of them may assume they sit next to the take, which is what broke all of them once.
Historical reference build: `projects/street-interview/` was private and gitignored. Its footage, approved take and colour target are optional development evidence, not inputs that a customer must re-fetch. Start with the bundled generic brand config, bind this project's assets and choices, and use zero-strength finishing unless a real colour reference is supplied.
## The five looks
One take, five finished treatments, all free. Each is a published grammar, not a palette:
| look | grammar |
|---|---|
| `clean` | title, two burned answers, a gold payoff, a dark end card |
| `subway` | SubwayTakes: a brand bar up the whole video, an occupation pill, sentence-case captions on a plate. The strongest "this is a series" signal |
| `karaoke` | TikTok-native: two or three words at a time, huge, centred, ~4 changes/sec |
| `bare` | no captions under the speech at all. The honest control |
| `doc` | documentary lower third, muted grade, no shouting |
## Cost
Each generation (12s, 720p, Seedance 2.0) is a paid call; `--fast` is the cheaper endpoint.
GooseWorks quotes the price. The five looks on top of a take cost nothing, which is why the
cheap axis of variation is the brand layer and not the generation.
Most of the development spend went on the assembled approach that this skill now exists to
stop anyone repeating.
## Quality checks
1. `check-cut.py` passes. TEN checks, A to I plus R, each one tied to something that was
actually rejected: format; length inside the 15s single-call cap; **provenance** (exactly one
generation manifest, and the render is the take plus one end card, so a stitched cut fails);
**prompt lint** (the clauses that were paid for are present, the vocabulary that was rejected
is absent, and the product is in shot 1 and the final shot); no caption across a cut;
graphics inside y=285..1635; **per-shot ambience floor** inside the real reference's band;
-14 LUFS +/-0.7 with true peak under -1.5 dBTP; every scripted line audible under Whisper;
and **R**, detail and black point on the render against the real-footage bands, measured at a
fixed 720px width so the number is not a function of the scaler.
A check that could not run prints `NOT RUN` and the gate exits non-zero. A WARN used to print
above `PASS` with exit 0 -- see Critical knowledge 19.
2. `check-cut.py --falsify` first. It plants an out-of-zone caption and asserts F reports it,
then an over-sharpened frame and asserts R reports it. A gate that has never failed is not a
gate, and five detectors written for this format were themselves the bug -- see `TAKES.md`.
3. `falsify-all.py` for the whole gate, not just those two. It plants a fault for every check
and reports which check caught it; anything that cannot fail is not a check. Free. Last full
run 2026-09-30: all eleven plants caught by their own check (A, B, C, C2, D, E, F, G, H, I,
R) -- and read Critical knowledge 18 before believing a MISSED, because two of the first
eleven plants were simply too weak.
4. `measure-pace.py` on the render. The gate does not look at pace. Reference band 1.54-1.62s.
5. Watched end to end (`/watch`). The gate cannot tell whether a face reads as AI or whether a
viewer can follow the video; every metric in an earlier scorecard passed while the cut was
incomprehensible. The gate now prints what it did NOT assess on every run, pass or fail,
because a PASS here has been read as "this is good" before.
6. Nobody has listened to the audio on this stack. Every audio judgement here is measured, and
that limit should be stated when handing a cut over.
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 |