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: strengthen image identity prompt anchors

rponeawac13698cb2bae
diff --git a/.agents/skills/hypit/references/playbooks/craft/image-prompt-style.md b/.agents/skills/hypit/references/playbooks/craft/image-prompt-style.md
index b171bb0e..1c36242f 100644
--- a/.agents/skills/hypit/references/playbooks/craft/image-prompt-style.md
+++ b/.agents/skills/hypit/references/playbooks/craft/image-prompt-style.md
@@ -37,10 +37,81 @@ Describe the final image directly. Do not write an edit-operation checklist, con
or long negative inventories containing unwanted concrete objects. Split mutually exclusive states
into separate prompts.
+## Keep the reality prefix on every photographic image
+
+The reality contract is an invariant prompt prefix, not a hint for the model. Every `gpt:Image`
+generation that depicts a photographic or photoreal scene must concatenate the exact prefix before
+its shot-specific body, including generations that also have one or more `gpt:Reference` children.
+References constrain identity, geometry, or continuity; they do not replace, shorten, or override the
+prefix. Keep the wording and order stable across a prompt family and change only the bracketed shot
+slot. A reference image with no prefix is a different visual contract and is not an acceptable
+shortcut. Non-photographic illustrations and deliberately graphic assets use their own plain visual
+contract instead of pretending to be iPhone footage.
+
+## A reusable description spine
+
+The strongest prompts turn adjectives into observable evidence and keep the same order every time:
+
+1. **Frame purpose and setting.** Name what the frame is doing (podcast, interview, product proof,
+ establishing view), where it is, and what part of the environment must remain visible.
+2. **Composition and camera geometry.** State shot size, subject position relative to the frame,
+ body lean, gaze direction, camera side and height, lens feel, and the background sectors or
+ landmarks that must be included. Relational instructions such as “subject slightly right,
+ leaning left, looking toward an unseen person on the left” are more reliable than three isolated
+ adjectives.
+3. **Identity and appearance anchors.** Establish the recurring person or product with concrete,
+ stable traits first (face shape, eyes, hair, skin, age presentation, distinctive marks), then
+ add body silhouette, wardrobe, accessories, and styling. For a male-presenting subject whose
+ physique matters, state a visible shoulder line — for example, “broad shoulders and a strong
+ upper-body silhouette” — and choose a half-body or wider frame that can actually show it. When
+ mixed Asian heritage is part of the intended identity, name the combination explicitly, such as
+ “Japanese-American mixed-race man” or “Korean-American mixed-race man”; do not leave a
+ reference-critical identity to the generic word “Asian.” Keep heritage as one stable identity
+ anchor, never as a bundle of stereotyped physical claims. Prefer a few renderable facts over a
+ stack of generic superlatives such as “extremely handsome.”
+4. **Action and physical relationships.** Give one readable starting action and posture. Say who the
+ subject is addressing, what supports their body, which hands are occupied, and how props contact
+ the scene. This explains a gaze, crossed legs, a microphone entering from frame left, or a person
+ leaning toward an off-screen interlocutor instead of leaving the model to invent the cause.
+5. **Props, branding, and scale.** Name object count, side of entry, relative size, and attachment to
+ the environment. Keep diegetic labels short and specify their placement; use supplied assets or
+ an explicit Track whenever lettering must be exact rather than trusting generated typography.
+6. **Light and material response.** Describe light source and direction, indoor/outdoor quality,
+ contrast, skin texture, surface texture, and the desired level of polish. “Natural indoor light,
+ fine skin texture, background still readable” gives the model physical evidence instead of only
+ saying “high quality.”
+7. **A small set of positive acceptance constraints.** Repeat only the few failure modes that matter
+ for this shot (for example, no depth-of-field blur, no broken/fragmented regions, background
+ visible). Avoid long negative inventories: they consume attention and often introduce the very
+ objects they name.
+8. **Reference roles.** For each ordered reference, say whether it locks identity, room geometry,
+ product shape, wardrobe, or UI. State what must remain stable and what the new shot is allowed to
+ change. Do not describe a reference-locked room a second time in prose as if it were a fresh set.
+
+For a prompt family, use this compact skeleton after the invariant prefix:
+
+```text
+Scene/purpose: …
+Composition/camera: …
+Subject/identity/wardrobe: …
+Action/posture/relationships: …
+Props/branding: …
+Lighting/material: …
+Acceptance constraints: …
+Reference roles: …
+```
+
+This is a writing order, not text to paste literally. Keep critical facts singular and unambiguous;
+use controlled repetition only for a fact that must survive across every view (identity, camera side,
+or lighting direction). Separate mutually exclusive poses, locations, or before/after states into
+different prompts rather than joining them with “or.”
+
## Keep people and cameras physically possible
- Make the first generated identity anchor a clear complete face with both eyes visible. Vary face
direction naturally only after that anchor exists.
+- If broad male shoulders are part of the design, keep both shoulder edges and enough upper torso in
+ the identity anchor; a tight face crop cannot establish that invariant for later references.
- Respect hand occupancy. A person holding a phone or product cannot simultaneously have both hands
free; assign only physically necessary tasks and avoid arbitrary left/right choreography.
- For a selfie, specify one real camera topology: an arm-held front camera, a mirror/reflective-surface
diff --git a/.agents/skills/hypit/references/playbooks/index.md b/.agents/skills/hypit/references/playbooks/index.md
index a36fc6fb..d73c76f3 100644
--- a/.agents/skills/hypit/references/playbooks/index.md
+++ b/.agents/skills/hypit/references/playbooks/index.md
@@ -53,6 +53,9 @@ file that some jobs never open.
- Write VLM instructions and image/video generation prompts in English. Keep original-language
dialogue, transcript, pronunciation, and quoted copy verbatim.
+- Every photographic `gpt:Image` prompt begins with the invariant reality contract in
+ `craft/image-prompt-style.md`, including when ordered `gpt:Reference` children are present;
+ references constrain the shot but never replace that prefix.
- Describe the desired visible state directly. Omit unwanted concrete objects instead of naming
them inside negations or hypotheticals that a model may materialize.
- Keep editorial captions, titles, stickers, cards, and callouts out of generation prompts. Author