Skip to content

armgit / hypit

Public Imported from GitHub
0

Clone any viral video with AI agents. Not just a script, the whole workflow: swap the face, the words, the B-roll, ship 100 variants in one command, and get your 100M views.

docs: name the defects that pass every structural check

rpone779652240dcd
diff --git a/.agents/skills/hypit/references/playbooks/craft/production-gates.md b/.agents/skills/hypit/references/playbooks/craft/production-gates.md
index 40aab9bb..d38486b1 100644
--- a/.agents/skills/hypit/references/playbooks/craft/production-gates.md
+++ b/.agents/skills/hypit/references/playbooks/craft/production-gates.md
@@ -81,6 +81,21 @@ visual QA.
## Gate 2: generate and review video takes
- Demand only the takes whose reference images have passed Gate 1.
+- **Before demanding any take with `generate-audio="true"`, trace its prompt back to the words it is
+ supposed to say.** Follow the edge from the take's `prompt=` to the `copy:Render` that produced it,
+ and confirm one of its `copy:Set` slots carries that Segment's `dialogue`. A prompt bound straight
+ to a `copy:Value` holding a performance direction is the defect this catches: the model receives no
+ words, invents its own, and the take sounds fluent and says nothing from the Script.
+
+ It survives every structural check. `hypit check`, `preview-check` and `plan` all pass, because the
+ graph is legal and complete — the prompt is a Text and the take consumes it. `whisperx:SemanticTake`
+ then fits the Segment's words onto whatever audio arrived rather than reporting that they differ, so
+ the Caption Track renders the Script over speech that shares none of it, and every downstream
+ measurement stays clean.
+
+ Two minutes of reading the Source here is the difference between finding it now and finding it in a
+ delivery whose every take has been paid for. `seedance-directing.md` and `../../seedance-kits.md`
+ name the slot the words travel through.
- Keep Seedance duration inside the selected model's declared range, read from that model rather than
from memory: the models differ and one accepts far longer takes. Invalid values fail validation rather
than being silently clamped.
@@ -109,6 +124,25 @@ visual QA.
- Inspect the durable Build and retrieve the final Artifact with `get`.
- Watch and listen to the complete delivery, including the first and last second. Verify dialogue,
captions, overlays, transitions, audio density, claims, and CTA as one program.
+- **Measure the delivery's own speech against the Script.** An agent reads pictures and cannot hear,
+ so "listen and verify dialogue" above resolves to nothing on its own, and a take whose spoken words
+ came from somewhere other than the Script passes `check`, `plan`, `preview-check` and both
+ measurements below without a mark. Transcribe what was delivered and read it against the words the
+ Script holds:
+
+ ```bash
+ hypit-reference-video-tools prepare_reference --video-path output/final.mp4 --observer agent
+ ```
+
+ Its `transcript` is measured locally by WhisperX from the delivery's own audio, so it is what the
+ video says rather than what the Source intended. Compare it line by line with the Script's
+ Segments. They correspond, or a take was generated from a prompt that did not carry its dialogue —
+ read `seedance-directing.md` and the Kit contract in `../../seedance-kits.md`, which name the slot
+ the words travel through.
+
+ The symptom this catches is total rather than subtle: the captions render the Script while the
+ voice says unrelated sentences, and alignment hides it by fitting the Script's words onto whatever
+ audio arrived. Run it before reporting any delivery that contains speech.
- Measure the delivery for empty picture before reporting it as finished. Watching finds a scene that
is wrong; it slides straight past a second of black between two shots, and a program assembled from
Selections can hold several. `ffmpeg -i final.mp4 -vf blackdetect=d=0.2:pic_th=0.90 -an -f null -`
diff --git a/.agents/skills/hypit/references/runtime.md b/.agents/skills/hypit/references/runtime.md
index bce57314..5d87d641 100644
--- a/.agents/skills/hypit/references/runtime.md
+++ b/.agents/skills/hypit/references/runtime.md
@@ -18,6 +18,36 @@ through the store rather than writing any secret into the file.
A capability whose credential this machine does not hold is still declared. Preflight names it before
any Build is submitted, which is the correct place for the author to find out.
+### Give the picture and video models room to run
+
+The generation models are what a Build waits on, and they are the one place concurrency is worth
+setting deliberately. Give Seedance and the image model **10 each**.
+
+On a Provider with per-model lanes — `@hypit/provider-kie` is the usual one — that is a lane apiece
+inside a pool wide enough to hold both, so neither starves the other:
+
+```json
+"kie.<project>": {
+ "use": "@hypit/provider-kie",
+ "config": {
+ "apiKey": { "store": "env", "key": "KIE_API_KEY" },
+ "defaultConcurrency": 20,
+ "laneConcurrency": { "seedance-2-mini": 10, "gpt-image-2": 10 }
+ }
+}
+```
+
+`defaultConcurrency` is the total pool shared by every lane, so it has to be at least the sum of the
+lanes or the lane numbers are a ceiling nothing reaches. Name the lane by the exact capability the
+Source reaches — `seedance-2-mini` and `gpt-image-2` above — since a lane key that matches no
+capability is silently inert. Read the Provider's README for the lane names it admits.
+
+The local Providers stay small: `media`, `whisperx` and `hyperframes` are bounded by this machine's
+cores rather than by a remote queue, and raising them buys contention.
+
+Set this while writing the Profile. Changing it later needs a Worker restart, and the restart rule
+below makes that costly once a Build is in flight.
+
## Execute the lifecycle
```bash
diff --git a/.agents/skills/hypit/references/vocabulary.md b/.agents/skills/hypit/references/vocabulary.md
index 267ce2a8..c5138cc6 100644
--- a/.agents/skills/hypit/references/vocabulary.md
+++ b/.agents/skills/hypit/references/vocabulary.md
@@ -51,10 +51,17 @@ will actually accept, which is what a mistaken attribute is checked against.
Use this decision order:
-1. Reuse one existing component when it fully expresses the requirement.
-2. Compose multiple existing components when their declared outputs and timing express it without
- changing what the element is.
-3. Declare a real vocabulary gap only when neither option works.
+1. Reuse one existing component when its declared vocabulary expresses the element exactly — every
+ property that changes what the viewer sees lands somewhere the package declares.
+2. Compose multiple existing components when their declared outputs and timing express it to that
+ same standard, without changing what the element is.
+3. Declare a real vocabulary gap when neither option works.
+
+**Resemblance does not qualify a package.** A tag that draws the same kind of thing, or that carries
+every property except one, leaves that element unexpressed, and an element the installed vocabulary
+cannot express as stated is a gap. Steps 1 and 2 apply where the declared vocabulary already carries
+the element as the evidence states it; step 3 is the ordinary outcome everywhere else, and
+`local-author-package.md` is the route for it.
Step 2 is not a way around step 3. Composition may express one element with several components; it
may not move an element out of its role because some other tag happens to draw the same shape. What