Most of us have a graveyard of saved articles, interviews, books, and courses. The material felt valuable when we found it, but it rarely returns when we face a real decision. Practical AI distillation tries to close that gap: instead of asking AI for another summary, we turn a useful way of thinking into a Skill that can be called on future problems.

In this article, “distillation” means knowledge extraction and workflow capture. It is an analogy, not neural-network training. Creating a Skill does not retrain ChatGPT or Codex and does not change model weights.

Two useful kinds of distillation

Person distillation asks how someone tends to think. The source set may include their public writing, books, talks, and interviews. The goal is not to copy catchphrases or impersonate the person. It is to find recurring questions, decision criteria, priorities, counterexamples, and boundaries. A responsible result says, “These sources suggest checking these factors first,” not, “This person would definitely say this.”

Content distillation asks what a book, course, or method can help us do. A writing book can become an article-review Skill. A product course can become a requirements-review Skill. Your own meeting practice can become a weekly-retrospective Skill. The source stops being something you must reread and becomes a tool that can act on a new situation.

Why use ChatGPT first and Codex second?

ChatGPT is a natural workspace for the first half: reading attached files, discussing ambiguous passages, comparing several sources, and exposing contradictions. Codex is well suited to the second half: writing the result into a maintainable directory, creating SKILL.md, organizing references and examples, and running repeatable checks.

OpenAI describes a Skill as a package of instructions, resources, and optional scripts that ChatGPT and Codex can follow as a reusable workflow. A Skill is therefore more than a very long prompt. It is a small method handbook that can be inspected, tested, versioned, and improved.

Step 1: choose a problem before choosing a famous person

“Distill this great thinker” is too broad. Prefer a target such as “review product features using the criteria in these public essays,” “edit technical articles using this book’s method,” or “turn this course into a weekly review process.” A narrow purpose makes it possible to decide whether the resulting Skill is actually useful.

Use material you own, are licensed to use, or can access publicly. If ChatGPT cannot read a video link, provide a transcript or your own notes. Do not bypass a paywall, and do not copy a protected book or paid course wholesale into the finished Skill.

Step 2: ask ChatGPT to produce a distillation pack

Attach the material in one chat or project, then use a prompt like this:

Extract an executable method from the material I provided. Do not produce a
generic summary. Deliver:
1. A source map explaining what each source supports.
2. Recurring principles and the evidence for each one.
3. The questions and decision sequence used when approaching a problem.
4. Applicable situations, non-applicable situations, and common misuse.
5. Positive and negative examples rewritten from the source material.
6. Contradictions, missing information, and any inference you made.

Label the source for every principle. Never present an inference as the
original author’s confirmed view.

Ask ChatGPT to export the result as five small files: source-map.md, principles.md, playbook.md, examples.md, and tests.md. Together they form the distillation pack. Keeping the source map separate matters: it lets you revisit a claim without burying the operational method under pages of quotations.

The tests should describe realistic inputs and the qualities of a good answer. Include at least one case where the method does not apply and one where the sources do not contain enough evidence. A Skill that cannot admit uncertainty will eventually invent certainty.

Step 3: ask Codex to build the Skill

Place the distillation pack in a new working directory and tell Codex:

Use $skill-creator to turn the distillation pack in this directory into a
reusable Skill. Read all five files before making changes.

- Define when the Skill should trigger and when it should not.
- Keep the stable workflow in SKILL.md and detailed material in references.
- Separate sourced facts, source-based inferences, and additional AI advice.
- Do not impersonate a person, invent views, or reproduce long passages.
- Run the cases in tests.md, revise failures, and verify again.
- Report the files created, installation method, and example invocation.

A lightweight Skill might look like this:

expert-review/
├── SKILL.md
├── references/
│   ├── source-map.md
│   └── principles.md
└── examples/
    └── review-cases.md

Do not trust it merely because the files exist. Test it with three ordinary questions and two deliberately awkward ones. Does it produce concrete steps? Can it show the basis for a recommendation? Does it say when the source material is silent? If it only copies tone and produces attractive quotations, you distilled a style, not a capability.

A complete content-distillation example

Imagine turning a writing book into a technical-article editor. ChatGPT first extracts checks such as: Who is the reader? What does the article promise? Is each important claim supported? Does every section move the argument forward? Which rules are firm, and which depend on genre?

Codex then turns those checks into a Skill with a fixed response pattern: restate the article’s goal, identify the three problems that most obstruct understanding, connect each problem to a sourced principle, and only then propose revisions that preserve the author’s voice.

Once installed, the everyday request becomes simple:

Use $technical-writing-reviewer to inspect this article.
Diagnose before rewriting, and connect every recommendation to a principle.

That is the real payoff of practical distillation. You do not get another summary waiting in a notes folder. You gain a repeatable perspective that can work on material it has never seen before.

Human judgment still belongs in the loop. People change their minds, methods have limits, and source collections are incomplete. Keep versions, add new evidence, correct mistakes, and preserve attribution. Distillation is not literally “stealing someone’s brain.” Done well, it is the respectful conversion of publicly available or legitimately obtained knowledge into a tool you can genuinely use.