If you do the same job with an AI agent more than a few times, stop rewriting the prompt and turn it into a skill.
A prompt is a moment in a conversation. A skill is a place where the procedure lives. Once the same instructions get pasted again and again, the wording drifts, a caution you included last time gets left out, and the output wobbles with it — and you rarely notice the omission.
I converted several parts of my blog operation: the pre-publication check, the post-publication verification, source checking, and turning AI news into article candidates for this site. This is what worked, what did not belong in a skill, and how small you can start.
Originally published in Japanese on June 18, 2026. This English version was written in September 2026.
Rewriting the prompt each time is less reproducible than it feels
For one-off work, prompts are fine. “Check this article.” “Summarize this news.” “Draft an outline for taupe.” Writing that fresh each time costs nothing.
Real operations are finer-grained. My pre-publication check alone covers title, description, canonical, OGP, headings, image alt text, internal links, external links, source strength, information that must not be published, and the site’s own voice.
Write that out by hand every time and something gets dropped — and a dropped check item is invisible by construction. That is the whole argument: fix the repeating work in place and make it editable afterwards, rather than getting better at composing prompts.
What the official skill formats actually specify
OpenAI’s Codex documentation describes skills as packaged instructions, resources, and optional scripts for reliably executing a workflow, following the open agent skills standard. The core is a SKILL.md with name and description in frontmatter followed by instructions, and a skill can carry scripts/ for deterministic behavior, references/ for documentation, and assets/ for templates.
The loading behavior is the interesting part. Per the same documentation, Codex initially sees only each skill’s name and description. The build-skills documentation puts a budget on that list — “at most 2% of the model’s context window, or 8,000 characters when the context window is unknown” — and Codex loads the full SKILL.md only once it selects that skill. The long procedure is not pasted in every time; it is fetched when it is needed.
Claude Code’s skills documentation describes a similar arrangement, with SKILL.md holding the main procedure and supporting templates, examples, scripts, and reference material alongside it. The implementations differ and I do not treat them as interchangeable — but the direction is the same: the agent goes and gets the procedure that fits the job, instead of being told everything up front.
Two pieces of context from OpenAI on where this is heading. Codex is being pushed beyond developers into analysis, marketing, sales, and creative work. And OpenAI announced on June 11, 2026 that it would acquire Ona — whose secure cloud execution environments are intended to help Codex take on longer-running work. Ona’s own site says the transaction closed on August 10, 2026, after the Japanese version of this article was published.
The longer an agent runs unattended, the less a clever one-shot prompt is worth, and the more you need the per-job procedure, verification conditions, prohibitions, the boundary where a human approves, and the recovery path when it fails.
What I put into skills
One is source-check. Another is article-seed, which turns AI news into candidate articles for this site. Beyond those: pre-publication check, post-publication sync, internal link insertion, and post-publication observation.
Nothing exotic. Source checking classifies each link — official announcement, primary source, community reaction, secondary article, weak source — and then classifies usability: usable as is, usable with context, needs further verification, better not used. That used to live in a prompt. Now it lives in a skill.
Article seeding is the same pattern. Instead of relaying AI news, the skill encodes the questions: what changes for the reader, can I actually try this myself, does it become a small experiment a reader could reproduce, are the sources sufficient, and does it leak internal information?
The result is that after a news pass, “give me the ones that could become articles” produces practical candidates for this site rather than news summaries. Dull, and disproportionately effective in daily use.
Checklists belong in skills; judgment calls do not
What converted well was checklist-shaped work: pre-publication checks, source checks, verifying the rendered HTML after publishing, finding internal link candidates, prioritizing article ideas, and organizing the metrics to review at 7, 14, and 28 days.
None of it needs to be reasoned from scratch each time — and reasoning from scratch is precisely how items get missed. In a skill, the order is fixed, the output shape is consistent, and last time’s lesson can be appended to the procedure.
What should not be converted: whether to publish, which phrasing to keep, how far to go when a source is weak, whether something is safe to disclose, and whether to execute an external post or an index submission.
OpenAI’s cookbook example on building an agent improvement loop with traces, evals, and Codex describes iterating with traces, feedback, and evals — and treats whether to run fully automatically or keep a human approval step as the developer’s decision. My blog operation lands on the same side: verification, organization, candidate generation, and diff review go to the agent; publishing, external sends, security judgment, and final voice stay with me.
Skills are not a way to hand everything over. They clear the repetitive work so the places a human must look are still visible.
Start with one checklist
You do not need a large automation to begin. A checklist is the easiest thing to convert, and the first version can be this small:
name: blog-publication-check
description: Before publishing, check title, description, headings, image alt
text, internal links, sources, and anything that must not be published.
Procedure:
1. Check that title and description match the article's content.
2. Check that the H2 flow reads for the intended reader.
3. Check that no image alt text is empty.
4. Check that there are at least one or two internal links.
5. Check that sources are official or primary where possible.
6. Check for secrets, account-specific values, and internal-only information.
7. Return findings split into: fix now, fix later, no action needed.Even at that size it beats retyping the prompt. Then split out reference files, add an output template, add a checking script, and add the items you actually missed in practice.
A skill does not have to be finished on day one; growing it in use is the natural shape. Mine grew exactly that way — something missed in a publication check, a hesitation during source checking, an article seed that came out sounding like a news summary, a repeated manual step in the post-publication docs sync. Each one moved into the procedure.
Grow the procedure, not the prompt
Prompts matter. Handling repeated work with a fresh prompt every time does not last.
Put it in a skill and the delegated scope becomes visible, the human decision points stay visible, the procedure can be corrected later, and it can be handed to the next agent. The skills I built are small, and the dullest work — pre-publication checks, source checks, seeding, post-publication sync — is where the payoff shows up soonest.
This is less about using AI than about growing the procedures your work already depends on.
Related
References
Codex and Claude Code differ in specification and implementation. They are cited here as nearby examples of the same idea, not as equivalents.