research

name	research
description	Deep creative research and conceptual exploration upstream of media creation. Develops rich, well-researched visual concepts before any pixels are made — for AI image gen, photography, illustration, concept art, or video. Covers: brainstorming visual ideas, mood boards, style/movement exploration, reference mapping, concept briefs, prompt packages, palette/composition analysis, and art history deep dives. Trigger when users want to: brainstorm visuals, explore a style, build a mood board, develop a concept brief, craft image prompts, plan a photoshoot, flesh out a vague idea, or when they say 'I want something that looks like X.'
metadata	{"tags":["concept","research","ideation","visual-language","prompt-crafting","mood-board","art-direction"],"related_skills":["genmedia"]}

Research

Great media start with great thinking. Move from a vague feeling to a concrete, actionable visual concept — then generate.

The six-step process and the pitfalls below are each their own ## H2. Pull the section you need with load_skill_section("research", "<slug>") rather than the whole file. Deep lookups (composition vocabulary, color harmonies, art movements, creative-strategy techniques, output templates) live in references/ — load on demand via load_skill_reference("research", "<name>").

When to use

• The user's brief is vague ("make something cool", "I want a vibe like X") and needs sharpening before generation.
• Brainstorming visual ideas, mood boards, style/movement exploration.
• Building a concept brief or a prompt package for a multi-image project.
• Reference mapping — finding specific artists / films / palettes that match a target.
• Art history or cross-cultural deep dives feeding into a project.
• Whenever the user names a style they want but you can't articulate what makes it work.

Not for: actual generation (do this first, then hand the brief to genmedia), pure technical questions (look up the relevant skill directly), one-shot prompt requests where the user is already specific.

How to use

Six-step process; each is a drillable section:

1. Listen and clarify — surface intent, destination, emotional target, references, constraints. See listen-and-clarify.
2. Check prior research — recall(query=...) (one call covers conversation content, session titles, and media prompts) first; then bash over $OKONOMI_HOME/research/. See check-prior-research.
3. Apply creative strategies — push beyond the obvious (conceptual blending, forced connections, assumption reversal). See apply-creative-strategies.
4. Research and explore — web_search / web_fetch for artists, movements, references. Pursue visual language, reference mapping, cross-pollination. For 2-4 genuinely independent strands (e.g. "films + comics + photography" for a noir brief), fan out via launch_agent instead of serial tool calls — cuts a 5-min serial session to ~2 min of parallel work. See research-and-explore for the pattern.
5. Synthesize and structure — pick a format (concept brief, prompt package, exploration document, photography direction). See synthesize-and-structure.
6. Save the research — write self-contained markdown to $OKONOMI_HOME/research/<slug>-concept-brief.md. See save-the-research.

Vocabulary lookups (composition, color, lighting, art movements, creative-strategy techniques, output templates) live in references/ — load on demand via load_skill_reference("research", "<name>").

Listen and clarify

Understand what the user is reaching for. Surface as needed:

Dimension	Question
Intent	What's it for?
Destination	Which medium? (Flux, photography, video, illustration)
Emotional target	What should the viewer feel?
References	What do they like? What do they not want?
Constraints	Aspect ratio, brand guidelines, medium limits

Infer what you can; ask about the rest. Bias toward asking rather than assuming.

Check prior research

Before starting fresh, search for existing work. Use the cross-session memory ladder (cheap → fall-back):

1. recall(query=...) — BM25-ranked FTS5 over every saved conversation, session title, and artifact. Hits carry snippets and goal/resolution bookends. The right starting point for "have we explored X before?".
2. Saved research files on disk:

   bash(command='ls $OKONOMI_HOME/research/')
   search(pattern='<keyword>', path='$OKONOMI_HOME/research/', file_glob='*.md')
   

   Read promising files to assess relevance.
3. recall(session_id=...) — drill into one promising hit for full metadata + artifacts + the conversation head/tail.

Try synonyms, related styles, artist names. If relevant prior work exists, surface it: "You explored something similar when you worked on X — want to build from that?"

Apply creative strategies

Push the concept beyond the obvious before research begins. Quick reference:

Strategy	Use when
Conceptual blending	Two domains can merge (Baroque + brutalism)
Replacement template	Symbolic substitution reveals hidden qualities
Forced connections	Need unexpected angles from distant domains
Assumption reversal	Flip fundamental premises to find new possibilities
Distance associations	Walk from obvious to surprising via semantic network
Perspective shifting	View through different lenses, scales, or timeframes

Use one or combine several. The goal: a direction that's specific and surprising. Full technique writeups in creative-strategies reference.

Research and explore

Let the concept's needs drive tool usage, not a fixed checklist. Err toward more research.

Tools

• web_search — artists, movements, techniques, cultural context, contemporary trends.
• web_fetch — full articles when search snippets aren't enough.
• bash(command="open <url>") or share image URLs in your reply — visual examples come back via okonomi's auto-attached multimodal pipeline so you can read them yourself.
• User-supplied files (mood boards, brand guidelines, prior briefs) — represent decisions already made and taste already established.

Threads to pursue (as relevant)

• Visual language — what makes a style work? Composition patterns, color relationships, texture, light behavior. Vocabulary in visual-language reference.
• Reference mapping — find specific artists, photographers, films, works. Describe what each reference contributes, not just "like this."
• Cross-pollination — unexpected connections across disciplines. A cyberpunk scene might learn from ukiyo-e composition.
• User context integration — existing materials anchor the search.

Parallel fan-out for independent strands — when the concept naturally splits into 2-4 independent angles (e.g. film cinematography + illustration lineage + contemporary photography for a "noir" brief), dispatch one sub-agent per strand instead of researching them serially. Cuts a 5-minute serial session to ~2 minutes of parallel work, and each strand gets full attention from its own agent context rather than the parent juggling angles.

Pattern — sub-agents write briefs to disk, parent reads + synthesizes:

import os
session_dir = get_session_dir()
strands = [
    ("films", "Research 1940s film noir cinematography — lighting, framing, B&W tonality. Cite 3-5 specific films with shot examples."),
    ("comics", "Research noir illustration / comic art lineage (Sin City, Eisner, etc.). Line-work conventions, panel composition."),
    ("photo", "Research contemporary noir photographers. Modern reinterpretations of the noir aesthetic."),
]
for slug, prompt in strands:
    out_path = f"{session_dir}/research_strand_{slug}.md"
    full_prompt = (
        f"{prompt}\n\n"
        f"Use web_search / web_fetch. Produce a tight 200-word brief. "
        f"Save it as markdown to: {out_path}\n"
        f"End by printing the file path so the parent knows where to find it."
    )
    launch_agent(task=full_prompt, description=f"noir/{slug}")

After launching, poll list_agents() until all show done. Then read the strand files, synthesize them into the final concept brief, and proceed to generation as usual. Sub-agents share the session dir, share the artifact registry, and auto-merge their tracked outputs back when they finish.

Don't fan out when the strands are interdependent (each depends on the previous strand's conclusion) — serial reasoning beats parallel context-loss for that case. Reach for fan-out only when the angles are genuinely independent.

Synthesize and structure

Pick the output format that fits. Full templates in output-formats reference:

Format	Use when
Concept brief	Complex or multi-image projects needing cohesive direction
Prompt package	AI image generation — a system for iteration, not just one prompt
Exploration document	Open-ended concepts with multiple possible directions
Photography direction	Planned shoots with shot lists and lighting direction

Save the research

Always save output as markdown to $OKONOMI_HOME/research/ with a descriptive filename (e.g., solarpunk-cityscape-concept-brief.md). The file must be self-contained — someone reading it cold should understand the concept, references, and direction without needing the conversation.

session_dir = get_session_dir()  # not used here, but available
import os
research_dir = os.path.expandvars("$OKONOMI_HOME/research")
os.makedirs(research_dir, exist_ok=True)
path = f"{research_dir}/<slug>-concept-brief.md"
with open(path, "w") as f:
    f.write(brief_markdown)
print(path)

Pitfalls

• Assuming instead of asking. The first description is often a placeholder for something deeper — running with it skips the real intent. Surface it with targeted questions.
• Asserting from memory. Relying on recall instead of a search produces confident errors, especially for non-Western traditions, contemporary artists, and niche movements. Verify first.
• Vague references. A confidently wrong artist / film / movement is worse than none. Name real things.
• Describing without showing. Words alone leave the concept ambiguous — use image search to surface visual examples alongside the writeup.
• Thinking in one-offs. A brief that describes a single image, not a visual system that could generate many, doesn't scale to the project.
• Generic over specific. "Nice warm lighting" tells the model nothing; "warm golden hour light raking across textured concrete at a low angle" does.
• Stopping at snippets. Search previews often aren't enough — fetch the full article; the depth is in the details.
• Stating the what without the why. "Use complementary colors" without explaining what it does to the viewer's eye gives the user nothing to reason with.
• Overriding the user's taste. Research and suggest — don't impose your own direction.
• Mis-scaling effort. A quick prompt needs one search; a visual world needs multi-threaded research across all tools. Match the depth to the ask.

References

Topic-deep material, available via load_skill_reference("research", "<name>"):

Reference	Contents
creative-strategies	Conceptual blending, replacement templates, forced connections, assumption reversal, distance associations, perspective shifting — full writeups with examples.
visual-language	Composition (rule of thirds, golden ratio, leading lines, negative space), color harmonies (mono, analogous, complementary, split-comp, triadic), lighting (hard/soft, Rembrandt, butterfly, split, rim), texture vocabulary, art-movement glossary (Impressionism → Vaporwave → Solarpunk).
output-formats	Templates for concept brief, prompt package, exploration document, photography direction.