Introduction
This post examines the Beautiful Article Skill — not an AI writing tool, but an engineering specification that teaches an Agent to run a complete magazine-style production workflow.
From "Prompt" to "Harness"
Instead of treating the Agent as a chatbot that answers questions, this Skill turns it into a "workshop supervisor" executing a full 8-phase workflow:
- Phase 0: Intake — decide whether to accept the job; if the user wants a dashboard instead of an article, stop and clarify.
- Phase 1: Source → Markdown — normalize any input (URL, PDF, DOCX, screenshots, text) into structured Markdown, with a 5-item checklist to prevent information loss.
- Phase 2: Editorial Planning — write a
plan.mdcontaining four sections: Brief (audience, article type, information retention ratio, tone), Outline, Theme, and Assets. - Phase 3: Plan Checkpoint — a hard gate. Five decisions must be confirmed item by item: article type, theme, layout, image mode, cover. The iron rule repeated throughout: "Never silently choose on the user's behalf." The Agent may recommend, but must not default.
- Phase 4: First Spread — hero screen + first section + one representative visual block, as scaffolding.
- Phase 5: Full Article Build — single Agent by default; long content isolated per section.
- Phase 6: Final Review — three perspectives: editorial, visual, technical.
- Phase 7: Repair — minimal-slice fixes; a
repair-log.mdis written only when repairs occur. - Phase 8: Delivery — final handoff and confirmation.
- 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
Each phase has defined inputs, outputs, QA methods, and decision nodes. This is not prompt engineering — it's workflow engineering.
Why "Never Silently Choose for the User" Is Top Priority
Most AI tools, when asked to write an article, just start writing with their own theme, structure, and tone. The Beautiful Article Skill designers consider this wrong. The workflow is built as a "toll-booth model": every key node must pause for user choice, and each decision item must be presented independently — never bundled into a single "all OK? yes/no" question.
The philosophy: AI's real value is not "doing it for you" but structuring your decisions. Users often don't know what they want at the start; clarity emerges through forced choices.
QA Protocol: When to Use SubAgents
The Skill distinguishes four QA modes instead of blindly spawning reviewers:
1. Inline self-check (5-item checklist) — Phases 1–2, where the main Agent's context is already warm; a SubAgent would be slower.
2. First Spread Reviewer SubAgent — Phase 4, since the first screen sets the tone and needs independent eyes.
3. Section Reviewer SubAgent — Phase 5, returning pass/fail by message instead of writing files; with 5–15 sections, nobody would read per-section review files.
4. Three-perspective Final Review SubAgent (writes final-review.md) — Phase 6, where the artifact is part of the deliverable and worth archiving.
This addresses the most common agent-workflow performance traps: over-spawning SubAgents and over-writing files. Differentiating QA per node is a mark of engineering maturity.
Theme System: Why AI Needs a "Design Signature"
The theme system is more than a color scheme — each theme defines a palette, fonts, typographic ratios, signature animations, and forbidden patterns. This counters the sameness of AI-generated pages (same rounded corners, same gradients, same spacing), giving every article a distinct visual personality.
Information Retention: Why the Default Is 100%
A subtle but important setting in the Brief: by default, 100% of source information is retained, producing a long-form web article. This is not a summarization tool — the assumption is that every piece of provided material has value. plan.md may mark "must keep" vs. "trimmable," but the default is full retention — a user-centric choice opposite to most AI writing tools' "default to concise" tendency.
Conclusion: Three Levels of Agent Workflow Engineering
1. Phase decomposition — explicit stages with inputs, outputs, and decision nodes. 2. Tiered QA protocol — choose QA intensity by value/cost ratio per node. 3. Decision framework — key decisions are never defaulted; they are surfaced so users clarify their needs through choosing.
Together, these make a production-ready harness rather than a demo-level prompt or toy-level chain.
Closing Thoughts
The Skill proves that when agent workflows are engineered seriously, AI can take over not just "writing" but the whole "making" process — from source intake to editorial decisions, visual design, implementation, review, and delivery. Humans shift from doing work in every step to making decisions at key checkpoints. You might answer fewer than 20 questions — "table or timeline for this section?" "generated image or screenshot?" "Bayer-style or Tufte-style cover?" — but each one shapes the final product. This isn't science fiction; it's what the Beautiful Article Skill already does.