Brand Gen Logo

Pi / Sage full-pipeline prompt

For Sage brand work in Pi, use the paste-ready prompt at docs/prompts/pi-sage-brand-gen-full-pipeline.md. It routes Pi through the typed brand_* tools, the brand-orchestrator subagent, exact-text gates, v2/DSPy review, GEPA-ready disagreement fields, and typed mutation loops. Keep this link instead of copying the full prompt into skill bodies.

Use this instead of the main brand-gen skill when the artifact is primarily a logo, wordmark, or lockup. Logo work has different constraints than standard brand materials.

How logo work differs

• Fewer refs, more silhouette discipline — logos need clean shape language, not reference-heavy compositions
• Geometric prompt language — say "flat vertical rectangles" not "columns"
• Tighter review loops — stop after each batch so the user chooses what equity to preserve
• Fragment locking — prompt fragments that work can be locked for all future generations
• Vector-friendly models — recraft-v4 and recraft-v4-svg are preferred over photo-realistic models

CLI and MCP tools

Logo has its own CLI and MCP server, separate from bgen:

• CLI: bgen <command> [opts]
• MCP tools: logo_generate, logo_feedback, logo_show, logo_compare, logo_evolve, logo_bootstrap, logo_inspire

Core workflow

1. Bootstrap (first use only)

bgen bootstrap

MCP: logo_bootstrap() — scans existing logo files into the manifest.

2. Gather inspiration

bgen inspire symbol          # browse logosystem.co by category
bgen inspire --url <url>     # custom inspiration URL
bgen inspire symbol --list   # list saved inspiration

MCP: logo_inspire(category="symbol") or logo_inspire(url="<url>")

Categories: symbol, wordmark, symbol-text, brown, beige, black, all.

3. Choose a mode

• `reference` — preserve recognizable structure from existing mark
• `inspiration` — net-new exploration, no structural constraints
• `hybrid` — preserve core equity while translating approved inspiration mechanics

4. Generate a batch

bgen generate \
  -p "Geometric pillar mark, three flat vertical rectangles ascending left to right, warm copper on transparent, clean silhouette" \
  -m recraft-v4 \
  --aspect-ratio 1:1 \
  --tag icon \
  --mode inspiration

MCP: logo_generate(prompt="...", model="recraft-v4", aspect_ratio="1:1", tag="icon", mode="inspiration")

Models for logo work:

Pass reference images with --reference-images /path/to/ref.png or --reference-dir /path/to/approved/.

5. Review, score, and lock fragments

bgen feedback v12 --score 4 --notes "Strong silhouette, simplify left column"
bgen feedback v12 --status favorite
bgen feedback v12 --lock "three ascending rectangles" "warm copper"

MCP: logo_feedback(version="v12", score=4, notes="...", lock_fragments=["three ascending rectangles"])

Locked fragments persist into all future prompts. Use them to protect equity that works.

6. Compare versions

bgen compare v10 v11 v12    # specific versions
bgen compare --favorites     # all favorites
bgen compare --top 5         # top 5 by score

MCP: logo_compare(versions=["v10","v11","v12"]) or logo_compare(favorites=true)

7. Evolve — learn from scored patterns

bgen evolve

MCP: logo_evolve() — analyzes prompt patterns across scored versions. Shows what works, what fails, locked fragments, and word frequency analysis.

8. Show manifest

bgen show              # full manifest
bgen show v12          # specific version detail
bgen show --favorites  # favorites only
bgen show --top 5      # top 5

MCP: logo_show() or logo_show(version="v12")

Scoring (same as main brand-gen)

5=near-ship, 4=good direction, 3=mixed, 2=weak, 1=reject.

The agent can propose scores based on conversation, but only persist when user intent is explicit. Always stop after each batch so the user clarifies what to keep and what to reject.

Shared brand context

Logo work benefits from the same brand foundation as standard materials. If a saved brand exists:

bgen show-session-summary --format json    # check current brand state
bgen show-identity --format json           # brand identity for tone/palette guidance
bgen ideate-messaging --format json        # messaging context for wordmark copy

Use the main brand-gen skill's identity and messaging to inform logo prompt language — tone words, palette direction, and shape language cues all apply.

What to avoid

• Never fabricate a lockup unless a real wordmark/lockup asset exists
• Don't overload prompts with detail — logos need clean, geometric language
• Don't skip review between batches — the user must choose what equity to preserve before iterating
• Don't use photo-realistic models for mark exploration — prefer recraft-v4 family
• Don't generate more than 3-5 versions before stopping for feedback