logo-explore-ai

Logo Explore (AI)

Goal: Batch-generate 25+ logo concepts as raster images using AI (Flux model), display in a comparison gallery, iterate on favorites.

Let: B := $HOME/.roxabi-vault/config/visual-charter.json P := <project-root>/brand/prompts/ O := <project-root>/brand/concepts/

Phase 1 — Context Discovery

1. Identify target project: $ARGUMENTS name/path → cwd (CLAUDE.md, README.md, package.json, pyproject.toml).

2. Check brand assets:

charter="$HOME/.roxabi-vault/config/visual-charter.json"
[ -f "$charter" ] && echo "CHARTER_FOUND" && cat "$charter" || echo "NO_CHARTER"

3. Research project identity: read CLAUDE.md, README.md, docs, configs — extract purpose, architecture, metaphors, personality.

4. Discover image generation environment:

for p in "$IMAGECLI_VENV" "$IMAGECLI_HOME/.venv" "$HOME/projects/imageCLI/.venv"; do
  [ -x "$p/bin/python3" ] && echo "VENV_FOUND: $p" && break
done
nvidia-smi --query-gpu=name,memory.free,memory.total --format=csv,noheader 2>/dev/null || echo "NO_GPU"
nvidia-smi --query-compute-apps=pid,process_name --format=csv,noheader,nounits 2>/dev/null

¬GPU ∨ ¬venv → abort with instructions. GPU conflict found → → DP(A) Stop conflicting process first | Proceed anyway Do NOT auto-kill.

Phase 2 — Concept Directions

Propose 5 directions (visual style + metaphor) via protocol: read ${CLAUDE_PLUGIN_ROOT}/../shared/references/decision-presentation.md (Pattern C). Ask to confirm/modify/add. Aim: 4-5 directions × ~5 variations = 20-25 concepts.

Phase 3 — Prompt Engineering

∀ direction: write 4-5 .md prompt files in P:

---
engine: flux2-klein
width: 1024
height: 1024
steps: 28
guidance: 4.5
negative_prompt: "blurry, low quality, watermark, ugly, text, letters, words, typography, font"
---

<prompt body — 3-6 sentences>

See ${CLAUDE_PLUGIN_ROOT}/examples/prompt.example.md for reference.

Rules: exact hex colors from charter; describe composition + background; reference render quality; state use case; ¬text/typography; vary angle/lighting/materials/detail per prompt. Filenames: 01-frosted-glass-hex.md, 02-frosted-glass-angled.md, etc.

Phase 4 — Batch Generation

mkdir -p "$brand_dir/concepts"
"$VENV_PATH/bin/python3" "${CLAUDE_PLUGIN_ROOT}/scripts/generate-batch.py" \
  "$brand_dir/prompts/" --output-dir "$brand_dir/concepts/"

Model loads once, generates sequentially (~20s each, int8 quantization). Failure → check nvidia-smi; reduce to 512×512; free GPU memory.

Phase 5 — Gallery

1. Read ${CLAUDE_PLUGIN_ROOT}/scripts/gallery-template.html.
2. Build IMAGES array from generated PNGs + prompt metadata.
3. Replace // __IMAGES_DATA__ with const IMAGES = [...]; optionally set const PROJECT_NAME.
4. Write to <project-root>/brand/concepts-gallery.html.
5. xdg-open "$brand_dir/concepts-gallery.html"

Phase 6 — Pick & Iterate

→ DP(A) Pick favorites (by number) | Generate more variations | Done

• Favorites + variations → new prompts (different angles/lighting/colors/materials) → re-run batch (existing PNGs skipped) → regenerate gallery.
• Done → summarize favorites; suggest: /logo-design for animated SVG, /logo-explore-svg for shape variations.

$ARGUMENTS