Logo Prompting

Writing a good logo prompt is a compression problem: you're translating a brand's strategy and voice into a short, literal, image-model-legible spec — while actively fighting the model's own priors, which default to AI-startup visual clichés (gradients, neural nodes, glowing orbs, soundwave bars) unless explicitly told not to.

This skill exists because a logo prompt has two failure modes, and they pull in opposite directions:

1. Too vague → the model falls back to generic SaaS-logo training-data averages.
2. Too literal / over-specified → the model draws exactly the clichéd icon you named (a phone for a calling app, a shield for security, a lightbulb for ideas) instead of an abstract mark that earns its meaning.

Good logo prompting threads between these: specific about constraints (palette, type, mood, what to avoid), abstract about the mark itself (suggest, don't depict).

Before writing a prompt: gather the four inputs

Don't start typing a prompt cold. Pull these four things first — they're the actual inputs, the prompt is just their compressed form:

1. Brand truth — read the project's brand.md if one exists (Strategy + Voice + Visual layers). If there's no brand.md, ask for the equivalent: what the brand is, what it's not, the archetype, the core promise. A logo prompt with no brand truth behind it is decoration, not identity.
2. The Visual layer specifically — Colors (exact hex, named tokens, and their usage rule — which color is primary vs. reserved-for-one-state), Typography (display face + weight + roman-only), Style keywords + reference brands. See references/brand-visual-layer.md.
3. Intake context — what is this mark actually for (wordmark, icon-only, favicon, both), what does it need to survive (16px favicon, dark mode, embroidery, print), who reads it (a shift worker at 2am glancing at a phone screen reads differently than a landing-page visitor). See references/logo-offer-questions.md for the fuller intake list — use it as an interview checklist when the user hasn't already supplied this.
4. The anti-slop guardrails — the specific clichés this brand must never look like (see references/anti-slop-discipline.md). Every brand has its own version of "don't look like a generic AI startup" — for a fintech it might be "no padlock icon," for a care-work app it's "no clinical cross," etc. Name the negatives explicitly; a model that isn't told what to avoid will reach for it.

Writing the prompt

Structure, in order:

1. What it is (wordmark / mark-only / lockup) and for what brand, one clause of category context.
2. Type treatment: exact face or face-description, weight, case, roman-only if that matters.
3. Color: exact hex or named token, and which color dominates vs. which is a reserved accent — don't let the model treat every brand color as equal-weight.
4. The motif, if there is one — described as an abstract suggestion ("a small integrated detail suggesting X"), never as a literal object unless the brand truly wants a literal object.
5. Negative space: the specific clichés to exclude (from references/anti-slop-discipline.md), stated plainly — "no gradients, no neural-network nodes, no glowing orb" reads better to Ideogram than a vague "make it not look like AI."
6. Mood / reference anchor: one sentence naming the feeling and, if useful, a real-world reference object (a radio dial, a carbon-copy ticket, a uniform patch) rather than a design-adjective salad.
7. A legibility/use-case constraint if relevant ("reads clearly at 16px favicon size").

One prompt, one direction. If the user wants multiple directions to choose from, write multiple structurally distinct prompts (different form, different type treatment, different motif — see the "structural variety over palette-swaps" rule below), not one prompt with several palette options bolted on.

Structural variety over palette-swaps

Borrowed from the Hallmark design skill's core insight: when producing more than one direction, each one needs its own locked token set (its own named colors, its own type pairing, its own motif) and its own form (wordmark vs. badge vs. mark-only vs. monospace-as-identity). Two directions that share a form and only swap the accent color are not two directions — they're one direction with a color picker. See examples/fleetline-four-directions.md for a worked example of four genuinely distinct directions from one brand brief.

Honest specificity, not adjective soup

"Modern, clean, professional, innovative" tells an image model nothing — every logo in its training data claims those words. Replace adjectives with things the model can actually render: an exact hex, an exact typeface name or type-genre, a named real-world reference object, an explicit exclusion list. If you catch yourself writing three unanchored adjectives in a row, stop and ask which one has a concrete visual referent.

The same trap has a sharper version one level up, at values rather than motifs: see moodboard-generator/references/panel-anatomy.md, "Depicting versus embodying". A padlock depicting security and a pair of hands depicting care fail the same way. Both draw the symbol of the claim instead of finding the visual choice that makes the claim true.

For retrieving the generated file afterwards (the permalink is not an image, and the file is WebP whatever you name it), see references/ideogram-asset-retrieval.md. Where a mark was specified against an exact ground colour or a photographic reference, the ground-colour and illustration-substitution checks in moodboard-generator/references/post-generation-checklist.md transfer; the panel-count checks do not.

When the ask is bigger than one mark

Sometimes the request is "help me get oriented on the whole visual identity" before locking a single logo — that's a moodboard job, not a logo prompt. Use the moodboard-generator skill instead, which owns the 3×3 grid template (Color Palette / Typography / Logo Exploration / Iconography / Photography & Graphics / Material Samples / Abstract Patterns / Mood Imagery / Application Mockup) and its own anti-slop discipline for that format.

After the fact: capture what worked

This skill is meant to get better as real logo prompts get run and judged. When a prompt produces a result the user actually likes (or explicitly corrects), that's signal worth keeping:

• Add or update a file under examples/<brand-or-style-family>.md with the brief, the final prompt, and one line on what worked or what had to change and why.
• If a pattern emerges across multiple brands (not just one brand's specific palette, but a reusable move — e.g. "monospace-as-primary-identity reads as 'record-keeping, non-digital' reliably"), promote it into this SKILL.md or into references/anti-slop-discipline.md as a named technique, not just left buried in one example file.
• Group examples by style family (e.g. analog-industrial.md, dark-mode-minimal.md, paper-record.md), not by client name alone — the point is to make the pattern reusable across brands, the way examples/fleetline-four-directions.md already demonstrates four such families from a single brief.

Reference files

• references/brand-visual-layer.md — how to read a brand.md Visual layer (Colors/Typography/Style) and turn its usage rules into prompt language.
• references/logo-offer-questions.md — the intake checklist for a logo job when no brief exists yet.
• references/anti-slop-discipline.md — the anti-AI-slop guardrails (no gradients, no neural-network/circuit-board/glowing-orb clichés, locked named tokens, structural variety), adapted from the Hallmark design skill for image-generation prompts specifically.
• examples/fleetline-four-directions.md — worked example: one fictional brand ("Fleetline," a voice-dispatch AI product), four structurally distinct logo directions from a single brief.

For a 3×3 moodboard/brand-exploration board, use the moodboard-generator skill instead of this one.