mono-color-skill 把印刷美學寫成 Agent 可檢查的規則

My reading of the README is that mono-color-skill is less an aesthetic than a spec. The name points to one ink, and the README's tagline covers both one-ink and controlled two-ink work, yet the README states that the skill defaults to controlled two-ink, with one dominant plate and one narrowly assigned accent plate, and that pure one-ink applies when a request explicitly asks for one-ink or monochrome. That default, and the numbers attached to it, are the part worth studying if you write skills for agents.

Adjectives like "restrained" or "editorial" give a model little to hold. This repo writes its visual system down as quantities, role assignments, and a refusal list, then has the agent inspect what it generates.

Defaults written as numbers

The ink rule reads like a config file. In two-ink work, the README says the dominant plate normally carries 70 to 85 percent of the printed area and the accent plate carries 15 to 30 percent. The accent is not free decoration: it "must have a specific job such as dates, annotations, selected objects, or overprint intersections." Two edge cases are closed in writing. The substrate is not a third color, and the darker color created where two plates overlap is not a third ink. Those sentences remove readings of "no more than two printing inks" that would otherwise stay open.

Empty space gets the same treatment. Visible empty paper occupies 25 to 55 percent of the canvas on an asymmetric editorial grid, and the README says it controls pacing. The composition step pairs that range with "one deliberate disruption," which reads to me as a budget for breaking the grid once, not an open invitation.

The palette is enumerated with hex values. The one-ink palette has eight inks, from Cobalt / Ultramarine at #2148B8 to Charcoal at #30343A, and a separate list gives nine two-ink recipes such as Cobalt + Terracotta and Cyan + Brick Red. Substrates are three paper tones, each with a list of the work it supports: Neutral White #FAFAF7 for crisp cultural, social, event, and image-led work; Cool Gray #E9E9E5 for architecture, technology, charcoal-led systems, and restrained branding; Pale Beige #F5F1E8 for tactile, travel, food, intimate, archival, or explicitly nostalgic subjects.

Naming what it will not make

The "Not this" section is the part I would borrow for my own skills. It lists outputs the skill rejects: a full-color photograph with a monochrome filter; a glossy mockup, 3D render, gradient poster, or cinematic scene; a centered template, card grid, sticker collage, or decorative blob system; marketing copy, invented branding, fake sponsors, URLs, or QR codes. To my mind, a negative list like this gives an agent concrete things to check against, where a positive description alone leaves the boundary vague.

One refusal targets a specific association. The README says the work is not automatically retro, yellowed, sepia, distressed, or nostalgic because it uses halftone or limited inks. Contemporary editorial is the default, and vintage treatment appears only when requested. Pale Beige paper stays available for the subjects the README lists.

Copying grammar, not pictures

The originality rule comes with a number as well. The README says the visual system was informed by a twelve-image research set, and it points to a references document with an image index, research notes, and attribution status. Rule 7 reads "References are grammar, not templates," and attaches a threshold: at least four structural variables change from every supplied reference. The refusal list adds that the output is not a reconstruction of a reference poster or artist signature. Even the repo's own generated examples are labeled as outputs, "not templates to reproduce."

Identity gets a counterweight. Rule 6 says supplied people, objects, and scenes remain recognizable. Rule 3 describes what happens to a photograph: it becomes dots, grain, clipped highlights, ink pooling, and mild registration drift between plates. Put together, a photo you supply should still read as its subject after that treatment, while the layout is rebuilt around the subject, intent, words, and image role.

The inspection step

The README describes five steps: read the input, choose the layout, assign the plates, compose the page, then generate and inspect. That final step names what to check: ink count, identity, hierarchy, texture, and originality. To my reading, ink count, identity, and originality each line up with a numbered rule: Rules 1, 6, and 7. The declared output is a generated raster image, the exact production prompt, and a short recipe. Having the prompt and recipe next to the image gives a reviewer something to compare against the rules.

Installation is a git clone into ~/.claude/skills/mono-color followed by a Claude Code restart, and the README says other agent environments can load SKILL.md as the skill entry point. If you maintain skills of your own, read this README beside one of yours and count how many of your instructions an agent could fail in a way you could detect.

GitHub: https://github.com/yanliudesign/mono-color-skill

Curated by Agent Palisade — practical AI for small and mid-sized businesses.