IP as Logo Skill for Hermes Agent: Minimal Mascot Guide

The IP as Logo skill turns a product brief into a set of minimal mascot candidates that still read as a character when shrunk to a 32 × 32 pixel favicon. It is for designers, indie developers, and maintainers who need a brand character rather than a full illustration, and who want the agent to propose directions, generate a batch, and hand back every result without filtering. If you already know the subject and just need one polished drawing, this is heavier than necessary.

What it does

The skill packages a complete workflow for producing a cute, highly simplified IP character: a compact symbol built from a handful of large shapes, two base colors, and one background color. The output is not a character sheet or a scene. It is a single square asset designed to survive extreme downscaling.

In practice the agent does four things. It gathers context about the product, proposes three directions, generates a batch of independent candidates once the user agrees, and delivers every candidate with a label, rationale, assigned corner, prompt and color mapping, and a path or URL. The skill is opinionated about restraint: it caps the shape count, forbids decorative detail, and treats generation as a random draw rather than a test to be passed.

You would reach for it when a repo, app, or community needs a recognizable face. You would not reach for it when you need a detailed illustration, a multi-pose character bible, or a logo that must satisfy a strict brand guideline document.

Before you start

This is an optional skill. It is not part of the default install and must be added on demand. The skill lives at optional-skills/creative/ip-as-logo, is version 1.0.0, and is MIT licensed. It runs on linux, macos, and windows.

The author is s1dashu, with the upstream project at s1dashu/ip-as-logo-skill; the Hermes port is maintained by Hermes Agent. The upstream tags are logo, mascot, branding, ip-character, image-generation, and creative.

The one hard prerequisite is an image-generation backend configured for the built-in image_generate tool. Check with hermes tools and enable a backend there if the tool reports none. The skill does not pick or switch models for you; the active backend is whatever you have configured.

Two other tools appear in the workflow but are conditional. vision_analyze is used only when the user explicitly asks for a compliance check. delegate_task is used only for large batches where the user wants speed. Neither is required for a normal run.

Because image_generate exposes no negative_prompt parameter, the skill always uses the main-prompt constraints mode: the natural-language Constraints: line stays inside the single prompt string. Modern instruction-following backends such as GPT Image, Seedream, FLUX, and Grok Imagine take the full prompt skeleton this way.

How the procedure runs

Discovery and context

The agent first parses the request for an explicit IP subject and any product context. It does not ask the user to choose a color mode unless the user explicitly wants to control it.

If no subject is given and the workspace is a product repository, the agent inspects read-only context before asking anything. It prefers the README, product docs, package or app metadata, landing-page copy, manifests, and design tokens. Context counts as sufficient when the product purpose, primary audience, and intended personality can be inferred with reasonable confidence.

When context is still insufficient, the agent asks one consolidated round of background questions: what the product does, who it serves, and how it should feel. There is no second questionnaire. The agent continues with the best supported interpretation after the answer.

The three-direction proposal

Once context is sufficient, the agent always presents three concise directions before generating, and explicitly proposes six independent candidates in one batch. It does not generate until the user agrees, unless the request already explicitly authorizes six outputs or asks the agent to proceed without another confirmation.

The three directions are chosen deliberately. If the user named a subject, the agent keeps that subject and proposes three distinct treatments based on silhouette, secondary color region, defining feature, or personality emphasis. If the user did not name a subject, the agent proposes three genuinely different subjects or metaphors, each tied to a different product attribute or brand promise. Three arbitrary animals with no rationale do not qualify.

Each direction is described in one compact line, and the proposal ends with a direct offer to generate six images using the distribution below. The discovery phase is not a long branding workshop unless the user asks for one.

Reading the user's response

The skill maps responses to batch shapes exactly:

If the user accepts all three directions and the six-image proposal, generate two independent variants per direction and label them A1, A2, B1, B2, C1, and C2. Assign A1, B1, and C1 to the lower-left and A2, B2, and C2 to the lower-right, so every direction is tested once from each side.
If the user selects one direction but accepts six images, generate six controlled variants of that direction labeled A1 through A6. Odd-numbered variants go to the lower-left, even-numbered to the lower-right.
If the user rejects the proposed quantity, directions, or distribution, follow the user's replacement instructions without arguing for the default.
For a pre-authorized reduced batch such as "generate exactly 2" where no directions were picked, prefer one direction with N variants labeled A1..AN, and state the direction and rationale in the report. Skip the three-direction proposal round whenever the request already authorizes a specific batch and forbids further confirmation. The always-present-three-directions rule applies only when a proposal round is possible.
For any other even default batch size, split candidates equally between lower-left and lower-right. For an odd batch, assign the extra candidate to either side deliberately and record the imbalance. Do not use bottom-center unless the user explicitly requests it.

Colors and generation path

Every candidate defaults to exactly three semantic colors in the complete image: two IP base colors plus one background color. The two IP colors are reused for facial marks rather than introducing more colors. An explicit user request for another count is followed.

If the user supplies a background palette, every supplied color is reserved for backgrounds unless they say otherwise. The agent chooses exactly two IP base colors independently for the subject and context unless the user also assigns subject colors. Historical or example palettes are never treated as a closed list of allowed backgrounds.

Before promising output, the agent confirms the generation path. In Hermes this is image_generate. If it reports no configured backend, the agent asks the user to enable one with hermes tools rather than fabricating results.

For large batches where the user wants speed, candidates can be parallelized with delegate_task, one candidate per task, all sharing the same product brief and constraints, each with an assigned direction or variant. Otherwise the agent makes separate image_generate calls in one turn, which keeps all results in the conversation.

One asset per call

Each subject is abstracted using the complexity budget below. Every candidate is generated as a separate full-resolution square asset. The agent never asks an image model to compose a contact sheet, grid, or multi-image sheet. Previous candidates are not used as image references when testing prompt-only reproducibility.

Delivery

Each batch is a one-pass creative draw. The agent generates every requested candidate once, then preserves and delivers every returned result as-is. It does not inspect outputs to block delivery, classify them as recommended or non-recommended, retry them automatically, or repair them with post-processing.

Every result is preserved and labeled. The final report includes each label, the IP direction and rationale, the assigned corner, the saved path, the prompt and color mapping, and dimensions. When the backend returns only a URL with no pixel size, the report says "backend-native square" rather than inspecting the image just to measure it. All results are presented together. Refinements or replacements happen only when the user explicitly asks for another draw.

The complexity budget

The budget is the heart of the skill. It is what keeps a mascot readable at 32 × 32.

Build one dominant continuous outer silhouette from roughly 4, 7 large basic geometric shapes. Merge or delete any shape that does not carry identity, expression, or recognition.
Use at most one species-defining feature, such as one large pouch beak, one pair of curled horns, or one broad visor. For limbless or featureless subjects like snakes, ghosts, or blobs, the defining feature can be a silhouette gesture such as one plump coil or one wavy hem, and the paired-features rule reduces to the two eyes.
Use at most two broad internal color regions corresponding to the two IP base colors. Keep the face to two eyes and, only when needed for expression, one tiny mouth. Omit eyebrows, highlights, nostrils, texture, outlines, and decorative marks unless essential for recognition.
Remove repeated feathers, scales, fur tufts, armor plates, buttons, screws, numbers, labels, and other illustrative detail.
Make simplification, cuteness, and an endearing baby-like personality the decisive qualities. Favor a large head, compact proportions, soft cheeks, widely spaced simple eyes, and a calm friendly expression when appropriate.
Require a readable black silhouette and recognizability at 32 × 32. If a feature disappears or becomes noise at that size, enlarge, merge, or remove it.

Shape language and composition

The composition rules exist to keep the character dominant and unambiguous.

Use thick, rounded, weighty contours and broad color masses.
Forbid sharp corners, pointed ears or beaks, needle-like tails, thin antennae, thin smiles, narrow gaps, and acute flame or feather tips. Replace every necessary tip with a visibly blunt rounded end.
Show both members of paired identifying features, such as ears, horns, wings, gills, or bells.
Show the character upright and emerging from the assigned lower-left or lower-right corner, filling about 85, 95% of the canvas so the IP remains visually dominant.
Cropping at the bottom or assigned side is welcome when it strengthens the sense of emerging from that corner, but do not prescribe exact edge contact or a fixed crop.
Never center or bottom-center the character unless the user explicitly requests it.
Preserve both members of paired identifying features within the visible composition.
Keep the artwork upright; never rotate the canvas or tilt the main mark without an explicit request.

Simplicity and visual treatment

Start from large, clean semantic shapes and the strongest possible simple silhouette. The character should be understood immediately, before any internal feature is noticed.
Prefer fewer, larger, softer forms over extra definition. Do not add a feature merely to explain anatomy or material.
Keep facial marks tiny, simple, and subordinate. Do not add glossy hotspots or detailed cavity rendering to eyes, mouths, noses, or other small features.
Keep the named background color visually solid and uniform, without scenery, texture, halo, vignette, or lighting variation.
Ask for the subtle dimensional effect only with the single sentence used in the Prompt skeleton. Do not expand it into numerical strength or instructions for gradients, highlights, or shadows. Incidental gradients, shading, or mild dimensionality returned by the generator are acceptable and must not trigger filtering or retrying.
Keep the requested visual direction graphic and simple rather than asking for clay, inflatable, plastic, plush, toy-like, or photorealistic rendering.

Color and canvas rules

Default to exactly three semantic colors in the complete image: two IP base colors plus one background color.
Choose the two IP colors from the product context, subject identity, intended personality, and user request. Organize both into broad purposeful masses; reuse one for facial marks and keep the other in one continuous defining region rather than scattering decorative fragments.
Choose both subject colors independently from the background. Favor clear, lively subject colors when appropriate, but do not impose global saturation, OKLCH, hue-shift, or chroma bands on the IP.
Choose the background freely for the context or from a user-supplied palette. Unless the user asks for vivid color, gently mute the background by lowering its saturation a little; keep it clearly chromatic and intentional rather than vivid, gray, or muddy. Historical palettes and examples are suggestions only, never an allowlist or mandatory default palette.
Preserve clear visual separation between the dominant IP silhouette, its facial marks, and the background. If a user-supplied background causes weak separation, adjust the subject colors first rather than replacing the requested background.
Across a batch, vary the two-IP-color strategies deliberately instead of repeating the same neutral-heavy combination.
Treat the two character colors as semantic color families. Incidental tonal variation within either family does not invalidate an output.
Name the intended solid background color directly. Ask for it to fill every open area and the unoccupied corners while the assigned emergence corner is occupied by the character. Do not use image-mode terms such as opaque, alpha, or transparency in the generation prompt.
Generate a direct 1:1 square with square outer corners (aspect_ratio="square" in image_generate). Accept the backend's native square resolution as-is; never resample merely to reach a specific pixel count.

The prompt skeleton

Describe the requested visual as an image only. Never tell the image generator that the image is a logo, brand mark, app icon, icon asset, or intended for any of those uses. Do not prepend use-case or asset-type scaffolding that reveals such a use. This rule applies only to the generation prompt; the surrounding user conversation and skill name may still describe the broader project.

image_generate takes a single prompt string, so always keep the natural-language Constraints: line inside the main prompt (there is no separate negative-prompt channel). Record the prompt and color mapping used per candidate in the generation report.

Create one complete full-bleed 1:1 square image.
Background: fill the entire square with solid <background>. Keep <background> visible in every open area and in the corners not occupied by the character; the assigned emergence corner must be occupied by the character.
Subject: place one extremely simplified, cute, endearing <subject> IP character on the background, reduced to one soft rounded continuous silhouette and one defining feature.
Complexity: use only 4–7 large basic shapes and at most two broad internal color regions. Use two simple eyes and add one tiny mouth only when it helps the expression. Remove every nonessential line, outline, anatomical detail, texture, and decoration. Keep the character readable at 32 × 32.
Color behavior: use exactly three semantic colors in the complete image: exactly two IP base colors plus the background color. Choose the two IP colors from the subject and context, organize both into broad purposeful masses, and reuse them for facial marks. Choose the background independently or follow the user's supplied background. Unless the user asks for vivid color, lower the background saturation slightly so it feels gently muted and restrained while remaining clearly chromatic, clean, and intentional rather than gray or muddy. Keep the IP, facial marks, and background clearly separated. Treat any example palette as optional inspiration, never as an allowlist.
Composition: keep the character upright and emerging from the assigned <lower-left or lower-right>, filling about 85–95% of the square so it remains visually dominant. Cropping at the bottom or assigned side is welcome when it strengthens the corner emergence. Preserve both paired identifying features. Never center or bottom-center the character.
Style: make simplification, cuteness, and lovable baby-like appeal the strongest qualities. Use large soft forms, compact proportions, thick rounded contours, and an ultra-clean graphic treatment. Prefer one clear shape over several explanatory details. Add an extremely, extremely subtle, almost imperceptible sense of depth through a barely-there neo-skeuomorphic treatment.
Finish: show only the character on the full-canvas background, with clean surfaces and normal square outer corners.
Constraints: Use no text or watermark. Add no borders, frames, cards, or presentation masks. Include one character only, with no extra subjects or scenery. Use no fragile lines, sharp tips, unnecessary outlines, tiny details, or decorative marks. Add no photorealistic material, dramatic bevel, glossy hotspot, deep occlusion, extrusion, strong three-dimensional rendering, or external cast shadow. Keep the background solid and uniform, with no texture, vignette, or lighting variation.

Delivery behavior

Treat generation as a stochastic draw, not a conformance test.
Generate the requested number of independent candidates once and deliver every returned image.
Do not inspect or report alpha, transparency, or background mode by default.
Do not block delivery, rank candidates as compliant or non-compliant, mark them as recommended or non-recommended, or automatically retry any result because of its background, colors, detail, composition, gradient, shading, or dimensionality.
Do not post-process a result to make it appear more compliant. If the user later requests another direction or replacement, generate a new independent candidate in response to that explicit request.

When not to use it

This skill is built for a single minimal mark, not for a full identity system. If you need a detailed character illustration, a multi-pose reference sheet, or a logo that must satisfy an external brand guideline, the constraints here will work against you. The source does not name a specific alternative skill for those jobs; it only notes that the skill's own rules are deliberately narrow.

If you already know exactly what you want and only need one image, the three-direction proposal and six-candidate batch are overhead. The procedure does allow a pre-authorized reduced batch, so state the count and the direction up front and skip the proposal round.

Limits and gotchas

The pitfalls section is short and worth memorizing.

Detail creep is the single biggest failure mode. Models add outlines, texture, extra colors, and scenery. The prompt skeleton's Constraints: line is load-bearing and must never be trimmed.
Naming the asset a "logo" or "icon" in the generation prompt triggers presentation framing such as badges, cards, and mockups. Keep the prompt purely pictorial.
Asking one call for a grid or contact sheet of variants produces small, inconsistent characters. One candidate per call, always.
If two candidates in a batch come back nearly identical, that is normal stochastic behavior. Deliver both; do not silently regenerate.

The delivery rules add their own limits. The agent does not inspect or report alpha, transparency, or background mode by default. It does not rank candidates, mark them recommended or non-recommended, or retry a result because of its background, colors, detail, composition, gradient, shading, or dimensionality. It does not post-process a result to make it appear more compliant. vision_analyze runs only when the user explicitly asks for a compliance check.

Finally, the verification checklist is the contract for a finished run: every delivered candidate has a label, direction rationale, assigned corner, prompt and color mapping, and file path or URL in the final report, and the batch count matches what the user approved, with no candidate withheld, retried, or post-processed.

What pairs with it

The source lists one related skill: pixel-art. It is a natural companion when the mascot needs to be rendered at small sizes in a pixel grid rather than as a smooth vector-style mark. Beyond that, the workflow leans on the built-in image_generate tool and, for explicit compliance checks, vision_analyze.

IP as Logo: Minimal Mascot Marks for Hermes Agent