When AI Learns to Make a Magazine: Engineering Agent Workflows Through the Beautiful Article Skill

1. The Journey of a Pile of "Scrap Paper"

Imagine you have a pile of raw materials — a web page link, a few PDF pages, some hastily jotted notes, or simply a bundle of files your boss tossed your way. Your task: turn them into something publishable that people will actually finish reading and want to share.

In the past, this took three people: an editor to select content, a designer to lay it out, and a front-end engineer to turn the design into a web page. Now, one person plus one Skill can do it all. And that person doesn't need to write code, do layout, or even write articles — they only need to know how to state requirements.

This is what the Beautiful Article Skill aims to do. It's not an AI tool that writes articles for you; it's an engineering specification that teaches an Agent to "make a magazine."

2. From "Prompt" to "Harness"

Many people treat an Agent as a clever chatbot: you ask, it answers, conversation over.

The designers of the Beautiful Article Skill clearly disagree. In this Skill's design, the Agent is not a question-answering tool but a "shop foreman" executing a complete workflow.

The workflow has 8 phases:

• Phase 0: Intake — Decide whether to take the job. If the user wants a dashboard instead of an article, stop and clarify.
• Phase 1: Source → Markdown — Convert any input (URL, PDF, DOCX, screenshots, text) into structured Markdown. A 5-item self-check ensures no information is lost.
• Phase 2: Editorial Planning — Form an editorial plan. Don't write HTML directly; first write a plan.md containing four sections: Brief (target readers, article type, information retention ratio, tone), Outline (structure, components needed per section), Theme (visual theme selection), and Assets (imagery strategy).
• Phase 3: Plan Checkpoint — A hard gate. Must stop and confirm 5 items individually: article type, theme, layout, image mode, cover. The designers repeatedly emphasize an iron rule: "Silently choosing on behalf of the user is forbidden." Every decision item must be listed separately and await the user's answer. The Agent may recommend, but must not default.
• Phase 4: First Spread — Hero section + first section + one representative visual block. A scaffold, so the user sees the direction.
• Phase 5: Full Article Build — Generate the complete article. Single-Agent by default; very long content isolated per section.
• Phase 6: Final Review — Three-perspective final audit: editorial, visual, technical.
• Phase 7: Repair — Minimal-slice fixes. A repair-log.md is written only when repairs occur.
• Phase 8: Delivery — Final handoff, confirming delivery decisions.

Eight phases, each with clear inputs, outputs, quality checks, and decision gates. This isn't prompt engineering — it's workflow engineering.

3. Why "No Silent User Substitution" Is the Highest Priority

One sentence appears at least three times in the Skill documentation: silently choosing for the user is forbidden.

What does that mean? In most AI tools, when you say "write me an article," the AI starts writing immediately — defaulting to its chosen theme, its decided structure, its assumed tone. If you're unhappy, you can only revise after the fact.

The Beautiful Article Skill's designers consider this approach wrong. They design the entire workflow as a "toll-booth mode": every key node must stop and let the user choose. What article type? Which theme? Wide or narrow layout? AI-generated images or screenshots? How should the cover be done?

Each decision item is an independent question — never bundled into a single "all OK? yes/no."

Behind this design philosophy lies a deep insight: AI's true value is not "finishing it for you" but "structuring your decisions." Users often don't know what they want at the start; needs become clear through forced choices. The Agent's job is to present those choices, not skip them.

4. Quality-Inspection Protocol: When to Spin Up SubAgents — and When Not To

Another subtle strength is how finely the quality-inspection protocol is segmented. Not every phase needs an independent SubAgent review. The designers explicitly distinguish four inspection modes:

• Main-Agent inline self-check (5-item checklist) — For Phases 1 and 2. The main Agent has already read the material; its context is hot. Spawning a SubAgent would only be slower.
• First Spread Reviewer SubAgent — For Phase 4. The first screen sets the tone; it needs independent eyes.
• Section Reviewer SubAgent (returns via message, no files) — For Phase 5. An article may have 5–15 sections; if each wrote a review file, no one would read them. A message-returned pass/fail suffices.
• Three-perspective final review SubAgent (writes final-review.md) — For Phase 6. Part of the deliverable; keeping an archive has value.

This design solves the most common performance trap in Agent workflows: over-spawning SubAgents and over-writing files. Not all quality checks need separate processes; not all reviews need archiving. Differentiating inspection intensity per node is a hallmark of engineering maturity.

5. Theme System: Why AI Needs a "Design Signature"

Another highlight is the theme system. It's not just a color scheme — it's a "design signature": each theme has a concrete palette, typography, layout ratios, signature animations, and forbidden patterns.

The reasoning: AI-generated content tends to look alike. The same problem plagues AI-generated web pages — identical rounded corners, identical gradients, identical spacing. By introducing the "design signature" concept, the Beautiful Article Skill gives every article a distinct visual personality, avoiding a homogeneous AI aesthetic.

6. Information Retention: Why the Default Is 100%

In the Brief section, there's a seemingly unremarkable setting: retain 100% of information by default, producing a long-form web article.

What does that mean? This Skill is not a summarizer. It will not delete your material in the name of "conciseness." Its assumption: you gave it this material because every piece has value. The Agent's job is to organize and present it better — not to decide which parts don't matter.

The plan.md will explicitly mark "must retain" versus "trimmable" items, but the default assumption is 100% retention. It's a user-centered design choice, opposite to most AI writing tools' "default to trimming" tendency.

7. Summary: Three Layers of Agent Workflow

The Beautiful Article Skill demonstrates the three layers a mature Agent workflow needs:

1. Phase decomposition. Split the workflow into explicit phases, each with inputs, outputs, and decision gates. 2. Quality-inspection protocol. Not every phase deserves the same level of inspection; choose by value/cost ratio. 3. Decision framework. Critical decisions must not be defaulted; they must be presented to users, letting needs crystallize through choice.

Together, these three layers make an Agent workflow that can actually land in production — not a demo-grade prompt, not a toy-grade chain, but a production-ready harness.

8. Closing Thoughts

The value of the Beautiful Article Skill is not that it lets an Agent produce a pretty article. Its value is proving one thing: when Agent workflows are engineered seriously enough, AI can take over not just "writing" but the entire "making" process — from material input to editorial decisions, from visual design to technical implementation, from quality review to final delivery.

This is a new collaboration model. Humans no longer do concrete work at every step; they make decisions at key nodes. The AI is no longer a passive tool but an active collaborator — one that reminds you: "This decision hasn't been made yet; I need your confirmation."

Perhaps future editorial work will look like this: you sit in front of a screen, and the Agent stops every few minutes to ask a question. "Table or timeline for this section?" "Generated image or screenshot?" "Bayer-style or Tufte-style cover?" You answer; it continues. Finally it delivers an article; you review; it repairs. Across the whole process, you make no more than 20 decisions — but each one defines the shape of the final product.

This isn't science fiction. It's what the Beautiful Article Skill is already doing.

References

• ConardLi, Garden Skills: beautiful-article, https://github.com/ConardLi/garden-skills
• easy-learn-ai project commit 3b57351: feat: add "Beautiful Article Skill" and related resources