aesthetic-research

Aesthetic Research

Summary

Researches a named aesthetic by collecting visual reference images, annotating them across the 7-dimension framework from aesthetic-literacy, and writing a structured knowledge profile to knowledge/aesthetics/<slug>.md. On repeat runs, appends a dated update section rather than overwriting.

When to Apply This Skill

Apply when:

• A named aesthetic needs visual evidence grounding before token production
• The user asks to "look up", "research", or "find references" for an aesthetic
• The agent is about to apply an aesthetic for which no knowledge profile yet exists
• The user explicitly requests an updated or deeper research pass

Do NOT apply when:

• A research profile already exists and the user wants to produce tokens/assets (use aesthetic-application directly)
• The aesthetic is well-covered in the dictionary and the user wants a quick characterization (use aesthetic-literacy alone)

Dependencies

Requires: aesthetic-literacy (7-dimension framework, slug conventions)

Knowledge Base Path

Before reading or writing any profile, resolve the knowledge base root:

Context	Resolved path
Project-level install	knowledge/aesthetics/ at the workspace root
User/global install (-g)	~/.agents/skills/knowledge/aesthetics/

Identify the workspace root by locating a parent directory containing .git, package.json, or .agents/. If the resolved path does not exist and you need to write, create it first.

Throughout this skill, knowledge/aesthetics/ refers to the resolved path above.

Workflow

Step 1 — Check for existing profile

Check whether knowledge/aesthetics/<slug>.md already exists.

• If yes: Note the existing image count and last updated date. This run will APPEND to the profile, not replace it.
• If no: This run creates the profile from scratch.

Slug format: lowercase-hyphenated (e.g., dark-academia, frutiger-aero, neubrutalism).

Alias check: If the user's requested term differs from the canonical slug (e.g., user said "webcore", profile slug is early-internet), record the requested term as an alias:

• If the profile already exists, append the new term to its aliases frontmatter field if not already present.
• If the profile is being created now, populate aliases with the requested term if it differs from the slug.
• Also add the term to the aliases field of the matching dictionary entry in the aesthetic-literacy skill (aesthetics/<slug>.md, co-located with that skill's SKILL.md) if that file exists and does not already list it.

  Note: If the aesthetic-literacy skill is installed at a different path (e.g., user-global at ~/.agents/skills/ vs. the current project install), the dictionary file may not be writable or locatable. In that case, skip the dictionary write and record the alias only in the profile frontmatter aliases field.

Step 2 — Construct query variants

Generate at least 2 query variants to broaden coverage:

• <aesthetic name> UI
• <aesthetic name> design
• <aesthetic name> web
• <aesthetic name> app (for digital-native aesthetics)
• <aesthetic name> graphic design (for print/historical aesthetics)

URL-encode the query: spaces → %20, special chars → percent-encoded.

Step 3 — Search for visual references

For each query variant, search using available image sources and browser tools. Good sources include Pinterest, Dribbble, Behance, Google Images, or any image repository with relevant design content. Use whatever tools are configured in your environment.

Step 4 — Save images

For each image found:

1. Save it to knowledge/aesthetics/<slug>/images/ with sequential naming: 01.png, 02.png, etc.
2. Note the source URL for reference

Target: 15 images
Minimum acceptable: 10 images (below this, flag evidence_level: limited)
Maximum: 25 images (stop here even if more are available)

If a result links to a page without a usable image, skip it.

Step 5 — Broaden search if needed

If total images collected across all query variants is < 10, broaden your search:

1. Try Dribbble: https://dribbble.com/search/<encoded-query>
2. Try Behance: https://www.behance.net/search/projects?search=<encoded-query>
3. Try a general image search

Even with < 10 images from all sources combined, always produce a profile — set evidence_level: limited in the frontmatter and note this clearly in the ## Research Updates section.

Step 6 — Annotate images

For each image, briefly note what you observe across the 7 dimensions. These are raw observations, not synthesis. Example:

Image 3: VHS scan lines, pastel pink/lavender gradient, katakana text, perspective grid floor, classical marble statue, dreamy unmoored composition

Step 7 — Synthesize dimensions

Aggregate annotations across all images. For each dimension, classify findings into three tiers:

• Canonical (≥70% of images): This property defines the aesthetic — deviation breaks it
• Common (30–70%): Frequently observed but not load-bearing — presence reinforces, absence doesn't break
• Variant (<30%): Occasional — signals aesthetic range, not requirement

Profile Format

Write to knowledge/aesthetics/<slug>.md using this exact structure. Field constraints:

• slug: lowercase letters, numbers, and hyphens only — must match the filename (e.g. dark-academia)
• image_count: integer ≥ 0, cumulative across all research runs
• evidence_level: standard (≥10 images) or limited (<10 images)
• new_aesthetic: boolean — true if not in the aesthetic-literacy dictionary
• aliases: list of alternative/community names that resolve to this slug

---
slug: <slug>
label: <Human-Readable Label>
first_researched: <YYYY-MM-DD>
last_updated: <YYYY-MM-DD>
source: <image-source>     # e.g. pinterest, dribbble, behance, mixed
image_count: <integer>
evidence_level: standard   # standard (≥10) | limited (<10)
new_aesthetic: false       # true = not in aesthetic-literacy dictionary
aliases: []                # alternative names / community terms that resolve to this aesthetic
---

# <Label>

## Dimension Synthesis

| Dimension | Canonical (≥70%) | Common (30–70%) | Variant (<30%) |
|---|---|---|---|
| Palette | ... | ... | ... |
| Type | ... | ... | ... |
| Texture | ... | ... | ... |
| Shape | ... | ... | ... |
| Motion | ... | ... | ... |
| Spatial | ... | ... | ... |
| Cultural markers | ... | ... | ... |

## Image Descriptions

1. [pin URL if available] — brief annotation across 7 dimensions
2. ...

## Analysis

_To be populated by the image-analysis skill._

## Connections

Related aesthetics from the dictionary: [list with one-line distinction from each]

## Research Updates

_Populated on repeat research runs._

Append-Only Rule

NEVER overwrite a profile. On repeat runs, add this section at the end:

## Research Update <YYYY-MM-DD>

- Additional images reviewed: <count>
- Evidence level after update: standard | limited
- Source(s): pinterest | dribbble | behance | other
- New canonical observations: ...
- Dimension changes: [dimension]: [was] → [now] (or "unchanged")

Do NOT modify earlier sections. The profile is an append-only research log.

Naming Novel Aesthetics

If the researched aesthetic is not present in the aesthetic-literacy dictionary, set new_aesthetic: true in the profile frontmatter. The profile becomes the canonical definition — there is no dictionary entry to fall back to.

Recording Aliases

When a requested term resolves to an existing aesthetic rather than introducing a new one, record the term so future lookups succeed without ambiguity:

1. Add the term to aliases in the knowledge profile frontmatter (knowledge/aesthetics/<slug>.md)
2. Add the term to aliases in the dictionary entry (aesthetics/<slug>.md, co-located with the aesthetic-literacy skill)
3. Do this even when no new visual research is performed — alias bookkeeping is a valid, minimal outcome of a research run.## Examples

Load EXAMPLES.md when uncertain about output format, ideal response structure, or to review anti-patterns.

Gotchas

• Never overwrite an existing profile. Profiles are append-only research logs. Overwriting destroys earlier dimension data, image counts, ## Analysis sections, and update history.
• Don't stop at fewer than 10 images without broadening first. Broaden to Dribbble, Behance, or general image search. Even if all sources combined yield < 10 images, produce the profile with evidence_level: limited. A limited-evidence profile is better than no profile.
• Keep image descriptions and dimension synthesis structurally separate. Raw per-image observations go in ## Image Descriptions. Frequency analysis goes in ## Dimension Synthesis. Conflating them makes the profile hard to update on future runs.
• Don't claim to identify exact font names from screenshots. Describe type category and visual characteristics instead (e.g., "geometric sans-serif, medium weight, wide tracking").