Download this Codex Skill when you need to turn sources and a reader question into a factual English guide, explainer, comparison, or review. It directs the article to answer the question immediately, connect checkable claims to their sources, keep details the reader can use, and apply a bundled Humanizer pass without changing the facts. Give it the exact question, audience, and source material; if the evidence is too thin, it should identify the gap instead of inventing an answer.
What to give the Skill
Start with the question the reader actually wants answered and who will use the answer. Add the source URLs or files the article should rely on, plus an existing draft if you want a revision. A writing sample can guide the voice. If you have a catalog of existing articles, provide it so the Skill can check whether another page already answers the same question.
The Skill can look for current authoritative sources when the task and available tools allow it. Source access depends on your Codex setup, so review the links and factual claims before publishing. Its default output is the finished article; ask separately if you want the internal claim audit or an edit log.
What to inspect in the finished article
| Check | What the Skill asks for | What you should see |
|---|---|---|
| Search intent | Identify the reader's exact question and answer it in the opening, with the condition that changes the answer. | The first paragraph gives a usable answer rather than an introduction that promises one later. |
| Evidence | Track each factual claim's source and distinguish documentation, observation, user-provided facts, and inference. | Decisive claims have a direct source link where readers need one; reported information is not presented as firsthand testing. |
| Useful detail | Keep topic-specific steps, criteria, failure signs, and observable results when the sources support them. | The reader can take a specific action or make a decision, and can tell what result to expect. |
| Humanizer pass | Remove formulaic AI prose, then compare the edit with the supported claims. | Names, numbers, citations, exceptions, and necessary qualifications still mean the same thing after the prose edit. |
Each paragraph should earn its place. The Skill asks the writer to remove text that would still fit an unrelated subject after changing only the subject name. It also tells the writer to stop when the evidence cannot support a useful article.
Install and try it
The ZIP contains one evidence-first-article-writer folder with five text files: SKILL.md, optional agents/openai.yaml display metadata, an MIT license for the original Skill, the complete Humanizer 3.0.0 reference, and that reference's separate MIT license. It has no executable script, account connection, or publishing command. The Humanizer reference uses a path inside the folder, so the Skill does not depend on a file from the publisher's computer.
Extract the folder into your repository's .agents/skills directory, or into $HOME/.agents/skills for your own Codex setup. Check that .agents/skills/evidence-first-article-writer/SKILL.md exists after extraction. The Codex Skills documentation describes both locations, explicit invocation, and what to do if an updated Skill does not appear.
Replace every bracketed field in this request with your own details. Give exact source URLs or file paths when you have them; if you need the Skill to find sources, say so explicitly and review what it finds.
$evidence-first-article-writer Write or revise a factual English article.
Reader question: [one precise question]
Audience: [who needs the answer]
Sources to verify: [exact URLs or file paths, or ask to find current authoritative sources]
Existing draft, if any: [paste text or say none]
Writing sample, if any: [paste text or say none]
Answer the question in the opening. Link only decisive claims to direct sources near their first relevant mention; avoid repeated links.
Keep useful steps, decision criteria, and limitations that the evidence supports.
Apply the Humanizer pass without changing facts, numbers, or qualifications.
If the evidence is insufficient, identify the essential gap. Return the finished article.Explicit $evidence-first-article-writer invocation is the clearest way to check that Codex found the Skill. Automatic selection depends on the task and the Skill description.
How we checked it
In one recorded Codex CLI run with v1.0.0, we gave the Skill a deliberately inaccurate technical draft and three official source pages. It produced an article that met eight checks written before the run, including a direct answer, supported claims, useful steps, and preserved qualifications. The test topic happened to be GitHub Actions artifact retention; this was only example content, with no required account or integration. Read the exact input, checks, unedited output, and editorial review if you want to assess this result. The v1.0.1 link-selection guidance was structurally checked but has not had a separate writing run. One example does not establish a general improvement in article quality or search performance.
Keep the facts through the final edit
The bundled Humanizer pass checks for staged openings, forced contrasts, repeated rhythms, inflated wording, and decorative formatting. After the rewrite, the Skill calls for a comparison with the supported claims: “may” must not become “will,” and an exception, number, citation, or meaningful detail must not disappear. Our editing guide gives more examples of fact-preserving prose edits.
