{
  "schema_version": 1,
  "skill_name": "analog-occultism",
  "evals": [
    {
      "id": "archive-portrait",
      "prompt": "Create a square portrait of a botanist examining a sealed specimen in analog occultism / industrial CRT noir.",
      "expected_output": "A concrete still-image prompt that preserves the person and specimen while using near-monochrome low-key lighting, one practical source, industrial framing, and restrained analog degradation.",
      "assertions": ["preserves the botanist and specimen as recognizable subjects", "specifies one dominant practical light", "uses near-monochrome crushed blacks and blown ivory highlights", "requires low-fidelity monochrome CRT defects such as soft raster, phosphor persistence, sync wobble, and dropped detail", "includes tactile analog signal treatment without excessive glitch", "avoids gore, glossy CGI, broad neon, and conventional horror"],
      "case_set": "dev"
    },
    {
      "id": "industrial-product",
      "prompt": "Prompt an image of a weathered field radio on a metal workbench in this aesthetic, with no readable brand or logo.",
      "expected_output": "A still prompt treating the radio as archival technical evidence with physical contact, cables, darkness, selective illumination, and original nonsemantic markings.",
      "assertions": ["grounds the radio on a workbench with plausible cables or contact", "uses a single screen or industrial practical as the light source", "keeps markings original and non-branded", "retains strong negative space and industrial geometry", "does not turn the scene into colorful cyberpunk or a nostalgic prop collage"],
      "case_set": "dev"
    },
    {
      "id": "reference-preservation",
      "prompt": "Use a supplied character reference to place the same person in a sealed archive chamber using analog occultism / industrial CRT noir.",
      "expected_output": "A reference-driven prompt that explicitly preserves identity and silhouette, adds the archive chamber treatment, and tells the image tool to use the actual reference.",
      "assertions": ["instructs the agent to pass the actual reference image to the generator", "separates PRESERVE from REINTERPRET or ADD instructions", "preserves identity and essential proportions", "adds one dominant practical source and severe contrast", "does not claim pixel-exact style transfer"],
      "case_set": "regression"
    },
    {
      "id": "slow-observer-video",
      "prompt": "Create a 10-second video of an inactive machine waking slightly in a dark archive, with quiet tension and no dialogue.",
      "expected_output": "A video shot prompt with one restrained mechanical action, slow automated-observer camera movement, stable practical lighting, sparse sound, and embedded CRT treatment.",
      "assertions": ["specifies a slow deliberate camera movement", "limits action to a restrained physical change", "keeps one dominant practical source stable", "specifies ambient sound and no dialogue", "avoids frenetic editing, whip pans, bright neon, and decorative glitch"],
      "case_set": "dev"
    },
    {
      "id": "video-keyframe-continuity",
      "prompt": "Plan an image-to-video or keyframe shot where an archival diagram slowly reveals one previously hidden node, then returns unresolved.",
      "expected_output": "A mode-aware video plan preserving composition, subject location, light, and visual grammar while making the reveal gradual and verifiable.",
      "assertions": ["chooses image-to-video or keyframes based on the need to preserve composition", "describes a gradual reveal and unresolved return", "preserves the dominant light and negative-space structure", "limits motion and signal effects to restrained changes", "includes bounded frame sampling and verification limitations"],
      "case_set": "release"
    }
  ]
}

# Image Prompting Reference

## Still-image structure

Lead with the subject and its spatial relationships, then specify the visual grammar:

```text
SUBJECT: [specific object/person/place/form], [position and relationship to frame]

LIGHT: one [screen/projector/industrial practical] from [direction]; selected
surfaces illuminated; surrounding space falls into deep shadow; clipped ivory
highlights and hard falloff.

AESTHETIC: analog occultism / industrial CRT noir; near-monochrome black, charcoal,
smoked gray, dirty silver, aged ivory; sparse midtones; tactile archival authenticity;
claustrophobic geometry; strong negative space; discovered technical artifact.

SIGNAL: this must look like a low-fidelity monochrome CRT screen capture or a
photograph of a CRT, not a sharp digital image with a filter. Use a small, soft,
interlaced broadcast raster; visible horizontal scanlines; coarse phosphor structure;
phosphor persistence trails; unstable vertical hold; slight horizontal tearing; sync
wobble; uneven brightness across the tube; clipped whites blooming around luminous
areas; crushed blacks hiding fine detail; dirty glass; dust; scratches; ghosting;
barrel distortion; and imperfect analog recording. Signal defects are structural and
must visibly reduce fidelity while preserving the essential subject.

MOOD: quiet unease, contemplation, precision, ritual, mechanical intelligence,
severe beauty, unresolved tension.

AVOID: broad neon color, glossy CGI, clean studio exposure, generic glitch art,
busy interface clutter, conventional horror, gore, cartoon treatment.
```

Use concrete nouns and verbs. For a portrait, specify what the light catches on the
face, hands, clothing, or reflective surfaces. For a product, specify contact with the
floor, table, cable, or enclosure. For a diagram or interface, specify what each panel,
line, or symbol communicates before describing the treatment.

## Capture-layer variants

Choose one before writing the prompt. Do not combine them accidentally.

### CRT-display view

The image is an image displayed on a square monochrome CRT. Emphasize the tube,
curved glass, phosphor, raster, scanlines, bloom, and signal instability. This is what
the previous test primarily achieved.

### Analog-camera documentary view

The subject is a real-world object or scene physically present in a room, photographed
or recorded directly by a period monochrome video camera. Say so plainly:

```text
A real physical [subject] in a real [location], captured directly by a period
monochrome analog video camera, not a digital render and not an image displayed on a
CRT. The camera sees actual surfaces, depth, reflections, lens focus falloff, exposure
imperfections, practical shadows, dust in the room, cable slack, worn edges, and
incidental background detail. Use a small low-resolution interlaced broadcast raster,
soft focus, scanlines, phosphor-like bloom, sync wobble, tracking noise, ghosting,
barrel distortion, crushed blacks, clipped whites, and dropped detail from the camera
and tape chain. The object must remain physically plausible and materially present.
```

A CRT monitor may appear in the room, but it is not required. The analog defects belong
to the camera, lens, tape, and playback chain. Do not describe only a digital scene
"with CRT effects," which encourages the model to render a clean synthetic object first.


Default to black, charcoal, smoked gray, dirty silver, aged ivory, and dirty white.
Use only a restrained cold green, blue-gray, or desaturated amber cast when it serves
the recording-medium illusion. A brief signal-error fringe is acceptable; broad neon
fields are not.

## Composition discipline

Prefer square or portrait crops, or tightly framed landscape shots when the brief
requires them. Anchor the frame with a luminous rectangle, receding grid, corridor,
foreground obstruction, or a small central object. Preserve large areas of darkness.
Controlled symmetry is useful, but add small physical imperfections so the frame feels
observed and recovered rather than digitally designed.

## Reference images

If a reference is supplied, pass it to the image generator. State what must be
preserved, what should be reinterpreted, and what must not carry over:

```text
PRESERVE: [identity, silhouette, proportions, essential markings, or composition]
REINTERPRET: [the supplied subject] through analog occultism / industrial CRT noir...
ADD: [new setting, light, or physical context]
DO NOT IMPORT: [unwanted logos, labels, background, or unrelated props]
```

The aesthetic is a treatment, not a license to alter the subject's identity. Use
`image-gen-reference` for the tool-specific edit workflow and anatomy gates.

## Text and symbols

Prefer no text unless it carries necessary meaning. When text is required, limit it to
short labels, indexes, specimen numbers, or system annotations. Use original symbols,
not recognizable logos or existing wordmarks. Verify visible text manually or with
vision inspection; generated text is not trustworthy merely because the prompt calls
it legible.

## Negative prompt

```text
bright even exposure, broad neon colors, RGB cyberpunk, vaporwave, teal-orange
cinematic grade, glossy commercial product photo, pristine CGI, clean vector art,
stock-photo smile, generic hacker room, random interface clutter, decorative glitch,
frenetic distortion, jump scare, gore, cartoon horror, fantasy ornament, nostalgic
prop collage
```

# Video Prompting Reference

## Shot template

```text
SHOT: [duration and framing]. [Subject] remains recognizable and grounded in the
space. One [screen/projector/industrial practical] is the dominant light source.

CAMERA: slow [push-in/pull-back/lateral track], stable mechanical movement, slight
servo hesitation, shallow focus breathing, no handheld energy.

ACTION: [one restrained physical action]. Allow long pauses and incomplete
information. Reveal [specific detail] gradually, then return to an unresolved state.

SURFACE AND SIGNAL: near-monochrome charcoal, smoked gray, dirty silver, and aged
ivory; crushed blacks; clipped highlights; restrained scanlines, interlacing, grain,
dust, scratches, halation, phosphor bloom, mild jitter, and occasional horizontal
interference embedded in the recording.

SOUND: [ambient machine hum / projector chatter / room tone]. No music unless
requested. No dialogue unless a visible speaker is specified.

AVOID: frenetic cuts, whip pans, action-camera movement, constant stimulation,
bright neon, decorative digital glitch, gore, jump scares, polished CGI.
```

## Motion rules

Use one primary motion and one or two secondary changes. Primary motion can be a
camera move, mechanical actuation, gradual reveal, or a slow change in focus. Secondary
changes can be screen flicker, exposure breathing, dust crossing the beam, or tracking
instability. Do not animate every object. Tension comes from duration, darkness,
partial evidence, and the anticipation of recognition.

A loop should not resolve the mystery. End with the subject returning to its initial
configuration, the light failing to reveal the key detail, or the observer stopping
just before recognition.

## Mode selection

- **Text to video:** use when the shot is conceptual and no opening composition must
  be preserved.
- **Image to video:** use when a supplied or generated still establishes the subject,
  framing, light, and surface treatment. Describe only the motion and temporal change
  that should occur.
- **Keyframes:** use when a transformation or state change must remain coherent. Keep
  the same subject location, dominant light, and negative-space structure between
  keyframes; vary only the intended state.
- **Continuation:** use when extending a shot. Preserve the established camera
  behavior, exposure, sound bed, and unresolved tension rather than introducing a new
  visual language.

For FLUX 3, read the `video-generation` skill and its prompting guide before invoking
any generation tool. Respect that skill's duration, frame-index, audio, and result
polling requirements.

## Video verification

Use FFmpeg or FFprobe to record container, stream, frame rate, duration, dimensions,
and audio details. Extract a bounded set of frames from the beginning, middle, motion
change, and end. Inspect those frames for:

- the subject remains recognizable;
- the dominant practical source remains dominant;
- blacks, highlights, and restrained tint remain coherent;
- movement is deliberate rather than frantic;
- signal treatment is embedded but does not obscure the subject;
- no accidental logos, text mutations, anatomy defects, color explosions, or digital
  corruption appear in sampled moments.

A sample validates sampled moments only. Record the sampling rule, timestamps, and
limitations in a concise inspection note, using the `ffmpeg` skill's video inspection
report when that workflow is active.