How to Turn a Decade of Writing into a Reusable AI Skill

Recently I tackled a problem more complex than simply asking AI to write articles: instead of refining a prompt, I first let AI systematically analyze a decade of my own writings and then organized the findings into a loadable writing /phodal-writer/ skill.

Why "write in my style" often fails

Most people try to make AI imitate their style by feeding a few sample articles and adding a short instruction. The output may look fine at the beginning but quickly drifts, reproducing only surface signals such as common words or tone. Deeper decisions—what topics to choose, how to open, how to structure sections, where to place judgment sentences, and how to conclude—are not captured, so the AI defaults to an averaged version of the author.

Style is not a set of isolated phrasing preferences; it is a hierarchy of writing decisions that evolve over time. Without explicitly extracting these decisions, AI can only mimic the few cues mentioned in the prompt.

Identifying style evolution

I divided my articles from the past ten years into three chronological blocks: 2014‑2019, 2020‑2023, and 2024‑2025. This segmentation prevents mixing distinct phase characteristics into a single average.

Analyzing the three periods revealed clear evolution:

Opening style: early articles start with a definition or direct experience; mid‑period pieces begin with a broad overview or framework; recent articles open with observation, practice, or version‑specific entry.

Structure progression: early works rely on lists; mid‑period shows stable hierarchical headings; recent works coordinate problem naming, layered decomposition, and engineering judgment.

Judgment sentence position: early pieces scatter judgments; mid‑period places them mainly at paragraph ends; recent pieces consistently place them at paragraph or section ends.

Paragraph density: early writing is fragmented with many short paragraphs; mid‑period paragraphs grow longer; recent writing features stable long paragraphs.

Ending strategy: early articles end with open reflections; mid‑period offers compressed but unstable conclusions; recent articles choose either a compressed judgment or an open invitation.

The table above shows that the baseline for my current style should be the 2024‑2025 period, while still preserving valuable traits from earlier phases.

From observation to rule

Extracting patterns is only the first step; the real work is converting observations into executable rules. For example, the observation "articles often start with a concrete observation" becomes the rule: "The opening must begin with a real‑world scenario, choosing one of three approaches (conflict, practice, or definition gap) and establishing the main question within the first three paragraphs." Similarly, "judgment sentences usually appear at paragraph ends" becomes: "Each major section must contain at least one strong concluding sentence placed at the end of a paragraph or section to lock in understanding."

Core rules derived

Problem progression: A main question is renamed and deepened across layers—phenomenon, engineering bottleneck, mechanism, and decision—guiding the argument forward.

Judgment sentence placement: Critical conclusions are placed at paragraph or section ends using stable cue phrases such as "In other words," or "This means," to compress information back to the core question.

Archetype first: Articles fall into a few stable archetypes (trend judgment, engineering methodology, project post‑mortem, protocol explanation, product release, personal reflection). Identifying the archetype determines opening, body decomposition, and closing decisions.

Quantitative anchors: Every key judgment is tied to a verifiable anchor—project name, version change, time span, code size, workload ratio—or at least a clear comparative context.

Skill organization

The resulting knowledge is stored in a directory structure that can be repeatedly loaded by AI:

<code>/phodal-writer/</code>
├── SKILL.md
└── references/
    ├── article_archetypes.md
    ├── writing_structure.md
    ├── style_core.md
    ├── quality_checklist.md
    ├── expression_patterns.md
    ├── revision_patterns.md
    ├── content_methodology.md
    ├── style_examples.md
    └── style_evolution/
        ├── 2014-2019.md
        ├── 2020-2023.md
        └── 2024-2025.md

When defining an article, the workflow is:

Read article_archetypes.md to select the appropriate archetype.

Consult writing_structure.md for the skeleton.

Use style_core.md to calibrate opening style and judgment‑sentence density.

After drafting, run quality_checklist.md for a four‑layer review of structure, style, content, and authorial voice.

This progressive disclosure ensures that each piece of information is available exactly when needed, reducing token waste and clarifying execution order.

Validation of the skill

The skill is not merely documentation; it must be able to generate a complete article that respects the defined structure, rhythm, and judgment placement. The current article itself serves as a test case, following the engineering‑methodology archetype: opening with a practical conflict, iteratively renaming the problem, extracting patterns, translating them into rules, and finally discussing the broader implication of experience migration.

Why this matters

Beyond teaching AI to mimic a personal style, the process demonstrates a generalizable method for encapsulating expertise into a reusable, loadable system. Prompt engineering solves a single‑round task, whereas a skill provides a repeatable framework that anyone can apply, amplifying the transferability of hard‑won experience.