Anycap Brand Mark Lab

AnyCap Brand Mark Lab

A repeatable pipeline for producing a production-ready brand mark with AnyCap image generation, distilled from two shipped projects (a GitHub org avatar derived from a parent brand's geometry, and a full logo redesign for a developer-tools collective).

The pipeline

brief ──▶ round 1: 4–6 composition candidates ──▶ no-text vision gate
  │              │                                        │
  │              ▼                                        ▼
  │      user rejects all? ──▶ round 2: NEW concept space (don't iterate on rejected elements)
  │                                                         │
  ▼                                                         ▼
i2i colorways on the winner (one variable at a time) ◀──────┘
  │
  ▼
64/44/32px decision board ──▶ final pick ──▶ delivery prep ──▶ provenance sidecar

1. Freeze the brief

One sentence of concept, required/forbidden elements (no text, no arrows, no reuse of existing logos), palette with exact hex stops, background, aspect (1:1, ≥1024px), and — most important — the surfaces where it will display. A GitHub org avatar lives at 32–44px; that constraint should shape every decision.

2. Generate candidates across the concept space, not around one idea

4–6 candidates, each a different concept (converging arcs, constellation, squircle container, pixel art, keycap, infinity loop…), not the same idea re-seeded. Varying only the seed wastes budget on one local optimum.

If the user rejects the whole round, do not iterate on the rejected elements. Explicitly jump to a disjoint concept space. Real example: after a round built on code brackets / circuit nodes / tools was rejected, the winning mark came from a deliberately disjoint round (pixel-art spark). Anchor prompt: "do NOT use any elements from the previous round".

3. Gate every candidate with vision read

anycap actions image-read --file <candidate.png> \
  --instruction "Any words, letters, numbers, watermarks, pseudo-text, or UI traces
   (borders/strokes around the icon background)? Answer CLEAN or HIT - <what>."

• Batch up to 10 files in one call.
• Icon-template borders count as HIT. A faint squircle stroke around the background is a UI trace; regenerate with "no border, no outline, no stroke — just clean flat background".
• Up to 10 images per call: include the composition question too ("does it keep the N-element ring?") when fidelity matters.

4. Explore colorways with i2i recolor — never re-roll from text

Once a composition wins, recolor it:

anycap image generate --mode image-to-image --param images=<winner.png> \
  --prompt "Recolor this exact logo mark, keep the composition and shapes 100%
   unchanged: replace the gradient with <hex A> to <hex B> … No text."

i2i preserves composition remarkably well. Change one variable at a time (hue family, mono vs color, dark vs light plate). State exact hex stops — #22D3EE → #2563EB follows orders; "blue-ish" doesn't.

5. Decide at real display sizes

Build a one-page HTML board: every candidate at 64/44/32px, plus the incumbent logo for contrast. 32px is the real canvas — a mark that only looks good at 1024px is not a logo. Watch for:

• multi-element marks fusing into mush (the reason the incumbent got replaced)
• thin vertical compositions wasting a square frame
• generic符号 (infinity loops, rockets, orbit atoms) — elegant but un-ownable; prefer marks with a unique texture (e.g. chunky pixels read crisp at any size and are instantly recognizable in a sea of flat gradients)

6. Delivery prep (the steps everyone forgets)

• Transparent corners. AI-generated "rounded-square on white" PNGs carry opaque white corners that show as an ugly plate on dark pages. Flood-fill the corner background to alpha before shipping.
• GitHub's camo proxy caches raw image URLs. After replacing an image in a repo, bump the README reference (logo.png?v=2) or the old pixels linger.
• Upload the avatar / apply the profile only as a separately authorized step — asset approval ≠ publish approval.

7. Keep provenance

A sidecar markdown per project: brief, references, candidate table with gate results, colorway list, final decision rationale. It makes the work reviewable, teachable, and reusable — and it is what turns a PNG into a case study.

Anti-patterns

• Re-rolling from text when only color must change (use i2i).
• Accepting a gate "CLEAN" from a previous round for a new edit — re-run the gate on every artifact.
• Asking the model for "a logo for X" with no concept direction — you'll get the industry's average logo.
• Judging candidates at full size only.