Imagine

Art-direction engine for text-to-image generation. Separates Style from Content to keep aesthetic identity consistent across an image series. Use when generating, rendering, or illustrating images (especially across a series), capturing the style of a reference image for reuse, or producing an image in a named/saved style. Not for captioning, image analysis for decisions (use ocas-look), or photo editing.

Imagine is an art-direction engine that treats image generation as a two-part process: Style Prompting and Content Prompting. Decoupling aesthetic DNA from subject matter lets a series of images share one visual identity while their content varies. This separation exists because style and content have different reuse patterns — style is stable across many images while content changes per request.

Interactive Menu

When invoked interactively, present a two-level menu. See references/interactive-menu.md for the full menu structure.

When to Use

• Create a series of images with a consistent visual identity (comics, storyboards, concept art).
• Translate a specific artistic style from a reference image into a reusable Style Prompt.
• Generate images using the current active text-to-image mode exposed by Hermes.
• Art-direct an LLM to produce a prompt that is style-pure (no content bleed).

When NOT to Use

• Analyzing user-provided images to drive downstream decisions → use ocas-look
• General web research on aesthetics or art history → use ocas-sift
• Editing or post-processing existing images → use an image editor
• Image captioning or description → use ocas-look

RULE — Read the Skill End-to-End Before Any Generation Run

Before ANY imagine run — especially the first after a gap in usage — read this SKILL.md and references/api_reference.md in full. Never reconstruct the methodology from memory of past runs and never guess the generation path: the skill documents the exact working procedure (current-model routing + GenTube browser bridge). If you catch yourself hand-rolling a substitute (hand-coded SVGs, a separate model, ad-hoc scripts), STOP — that is the signature of not having read the skill; fix the blocker instead. (Confirmed 2026-09-19: a banner run skipped the full read, hand-coded SVGs for ~50 turns, then found the documented path.)

Responsibility Boundary

Imagine does: style library management, style extraction from reference images, art-directed prompt synthesis, image generation through the current active conversation model when that model is text-to-image capable, and journaling of every generation and extraction run.

Imagine does not: analyze user-provided images to drive downstream decisions (that is ocas-look), perform general web research on aesthetics (that is ocas-sift), edit/post-process existing images, call the separate image_generate tool, or bypass the active model by calling Pollinations/Pollination/FAL/direct provider endpoints.

Adjacent responsibility: ocas-look consumes images as decision inputs; Imagine produces images as creative output. If a request is "look at this image and do X with the info," route to Look; if it is "make an image that looks like this," route to Imagine.

Optional Skill Cooperation

Imagine functions standalone. When present it may cooperate with:

• ocas-sift — research on a referenced artist, movement, or visual vocabulary before extraction.
• ocas-weave — resolving references to known people (portrait subjects) for identity context.

Imagine never depends on these skills and must run normally if they are absent.

Journal Outputs

• Action Journal — emitted by imagine.generate and imagine.style.save (external HTTP side effect or persistent state write).
• Observation Journal — emitted by imagine.extract and imagine.library.list (analysis or read-only enumeration, no side effects).

Every run produces exactly one journal file. See references/journal.md for the record schema.

Core Methodology

See references/style_prompt_guide.md for the full Style-Content separation methodology, the five standard style sections, and the overspecification rule.

Operational Flows

Flow 1: Image Generation (existing style)

Use when the user wants an image in a known or predefined style.

Checklist:

• Style selected from default_styles.md or styles.jsonl
• Content Prompt describes only what is in the scene (no style/color/lighting keywords)
• Final prompt = Style Prompt + Content Prompt (style first)
• API call executed successfully
• Action Journal record written with final prompt + image URL

1. Select Style: retrieve a style definition from references/default_styles.md or a previously saved Style Prompt in {agent_root}/commons/data/ocas-imagine/styles.jsonl.
2. Expand Content: turn the user's subject request into a detailed Content Prompt.
   • Constraint: describe what is in the scene, never how it looks (no colors, no style keywords, no lighting direction).
3. Synthesis: concatenate Style + Content into the final API call.
4. Execute: submit the synthesized prompt to the current active conversation model when it is text-to-image capable. Do not call the separate image_generate tool. Do not call Pollinations/Pollination, FAL, or any direct provider endpoint unless explicitly asks for that backend.
5. Journal: write an Action Journal record with the final synthesized prompt and the resulting image URL.

I/O Example:

• Input: imagine.generate --style noir --content "a detective standing under a streetlight at night"
• Output: {"status": "ok", "image": "<active-mode image URL or file path>", "journal_id": "2026-06-27_abc123"}

Concise art-direction output schema: Art-direction output is formatted as structured JSON containing prompt text, style weights, and negative prompts:

json

{
  "prompt": "<Style Prompt + Content Prompt concatenated, style first>",
  "style_slug": "noir",
  "style_weights": {"mood": 0.8, "composition": 0.7, "color": 0.9, "texture": 0.6},
  "negative_prompt": "blurry, low quality, distorted, watermark",
  "content_summary": "a detective standing under a streetlight at night"
}

Rules: style_weights are normalized 0.0–1.0 relative weights per standard style section; negative_prompt carries the style library's default exclusions unless overridden; prompt is always the executable style+content concatenation (per Flow 1 step 3). Use this schema for any art-direction output consumed programmatically.

Flow 2: Style Extraction (image → style)

Use when the user provides an image and wants to capture its soul for future use.

Checklist:

• Reference image loaded (path or URL)
• Multi-modal analysis completed
• Style extracted in exhaustive detail (no object/content names)
• Organized into five standard sections
• Style Test image generated and verified (no content bleed)
• Style Prompt saved to styles.jsonl
• Observation Journal record written

1. Visual Analysis: use a multi-modal LLM (via vision_analyze or equivalent) to analyze the reference image.
2. Exhaustive Extraction: apply the extraction prompt in references/style_prompt_guide.md.
   • Requirement: describe the style in exhaustive detail without naming any object or content.
3. Semantic Organization: organize the raw description into the five standard sections:
   • Perspective & Composition
   • Lighting & Shadow
   • Color Palette
   • Brushwork & Technique
   • Image Framing & Balance
4. Verification: generate a Style Test image of an unrelated, simple subject to confirm the prompt is robust and free of content bleed.
5. Save: append the resulting Style Prompt to {agent_root}/commons/data/ocas-imagine/styles.jsonl.
6. Journal: write an Observation Journal record for the extraction.

I/O Example:

• Input: imagine.extract --image https://example.com/ref-photo.jpg
• Output: {"status": "ok", "style_name": "extracted-noir", "style_prompt": "...", "test_image_url": "...", "journal_id": "2026-06-27_def456"}

Commands

• imagine.generate --style <name|prompt> --content <description> — generate an image in a specific style.
• imagine.extract --image <path|url> — analyze an image and produce a structured Style Prompt.
• imagine.library.list — list all available predefined and saved styles.
• imagine.style.save --name <name> --prompt <prompt> — save a new custom style to the library.
• imagine.journal — read the most recent run record (final prompt + resulting image URL).

Recovery Behavior

This skill implements the recovery contract from spec-ocas-recovery.md.

• Evidence: Every generation/extraction run writes an evidence record to {agent_root}/commons/data/ocas-imagine/evidence.jsonl, including no-op runs. The not_activity_reason field is mandatory when no side effects occur.
• Gap detection: Not applicable — on-demand only.
• Degraded mode: When the current active conversation model cannot emit image attachments through the current surface, logs degraded: current_model_t2i_unavailable and returns the interface error without using fallback backends.
• Log compaction: Evidence and history logs older than 30 days compacted. Last 7 days retained.

Storage Layout

text

{agent_root}/commons/data/ocas-imagine/
  config.json        # ConfigBase fields + Imagine defaults
  styles.jsonl       # Custom user-created style prompts (append-only)
  history.jsonl      # Append-only log of generation prompts and results
  intents.jsonl      # Append-only log of user intents per run
  evidence.jsonl     # Append-only evidence records (recovery contract)
{agent_root}/commons/journals/ocas-imagine/
  YYYY-MM-DD/
    {run_id}.json    # One journal file per run

Config follows ConfigBase from spec-ocas-shared-schemas.md. All paths use the {agent_root}/commons/ root per spec-ocas-storage-conventions.md.

Implementation Details

See references/api_reference.md for the current-model execution contract and failure handling.

Background Tasks

Imagine has no operational background tasks. The universal imagine:update self-update cron is standard and is not enumerated here.

Self-Update

See references/self-update-imagine.md.

Visibility

Public.

Gotchas

• Content bleed invalidates a Style Prompt — If a Style Prompt references specific objects, people, or scene content, it will produce inconsistent results across subjects. Always verify with a Style Test image of an unrelated, simple subject before saving.
• On-demand only — no background tasks — Imagine has no scheduled cron jobs or heartbeat tasks. It runs purely on invocation. A generation failure won't be retried automatically.
• Style-content concatenation order matters — The final API prompt must be Style Prompt first, then Content Prompt. Reversing the order causes the model to prioritize subject over aesthetics.
• Current-model failures are terminal — When the active conversation model cannot emit image bytes/attachments through the current surface, there is no built-in retry in the skill itself. Log degraded mode and report the interface problem. Do not route around it via image_generate, Pollinations/Pollination, FAL, or direct provider calls.
• Validation triple is mandatory — Every generation must produce entries in history.jsonl, a journal file, AND evidence.jsonl. A generation missing any of these is considered invalid per the OKR data_integrity target.
• Skipping the full read is the #1 failure mode — Every usage regression traces to not reading SKILL.md + references/api_reference.md end-to-end before generating. Read first, generate second. Hand-coded SVG substitutes for image generation are prohibited when the documented path is available.

Error Handling

Failure	Handling
Current model image path unavailable (interface/tooling error)	Log degraded: current_model_t2i_unavailable to evidence.jsonl, return the interface error to the user, and do not use fallback backends
vision_analyze fails on reference image	Retry once with explicit prompt asking for visual style description; if still failing, report error to user with image format requirements
styles.jsonl is corrupted or unreadable	Initialize fresh styles.jsonl with only built-in defaults from references/default_styles.md; log corruption event to evidence.jsonl
Content Prompt accidentally contains style keywords	Halt generation, report content bleed to user, request pure content description
Style Test image shows content bleed	Do NOT save the Style Prompt; report extraction failure, suggest providing a cleaner reference image
Journal write fails (permissions/disk)	Log failure to stderr, still return generation result to user, flag evidence record with journal_write_failed: true
Invalid image path or URL in extract command	Return error with accepted formats: local file path, http(s) URL; do not attempt download of unsupported schemes

Support File Map

File	When to read
references/style_prompt_guide.md	During Flow 2 step 2 — when extracting style from a reference image
references/default_styles.md	During Flow 1 step 1 — before picking a style, to browse available presets
references/soma.md	When user requests or you identify a soft luminous gradient aesthetic
references/noir.md	When user requests or you identify a flat silhouette + low sun aesthetic
references/hiro.md	When user requests or you identify a woodblock print aesthetic
references/comic.md	When user requests or you identify an ink linework + warm ground aesthetic
references/candy.md	When user requests or you identify a plein-air + opaque color field aesthetic
references/vaporware.md	When user requests or you identify a retro consumer electronics aesthetic
references/api_reference.md	When executing generation; contains the current-model routing contract
references/journal.md	Before writing any journal file; contains the record schema

Validation Rules

• Every generation writes to both history.jsonl and a journal file. A generation with no journal is invalid.
• Style Prompts must not reference specific objects, people, or scene content.
• Content Prompts must not reference colors, lighting, materials, or style keywords already covered by the selected Style Prompt.
• All filesystem writes stay within {agent_root}/commons/data/ocas-imagine/ and {agent_root}/commons/journals/ocas-imagine/.

OKRs

• schedule_adherence: 100% — on-demand only; every invocation completes or reports error within timeout. Measured via evidence.jsonl timestamps.
• data_integrity: 100% — every run produces valid history.jsonl + journal + evidence.jsonl records. A run missing any record is invalid.