content-refresher

---
name: content-refresher
description: Decides whether a slipped page should be refreshed, consolidated or left alone, then writes the exact edits that win the ranking back. Use when an old page has lost its traffic.
---

# Update an old page and win the ranking back

A verdict, the exact edits, and the rewritten sections ready to paste in.

## Claude does

1. **Ask for {PAGE_URL} and {TARGET_QUERY}. Refuse to start without both.** Ask the position now and a year ago; with no rank data, say so and work from today's SERP. Never guess a ranking or a traffic figure.
2. **Read the live page and the current top three results.** What those three answer and this page does not is the whole job.
3. **Issue one verdict**, on the SE Ranking triggers: KEEP (top three already), REFRESH (slipped, still the best asset), CONSOLIDATE (a second page owns the query; merge, redirect the weaker), KILL (query dead or never relevant). One sentence saying which.
4. **List the decay causes actually found**, never generic: stale dates and figures, sub-questions the top three answer, a title link nobody types, sections all three carry and this lacks.
5. **Write fixes as edits, not advice.** Each names the section, the current text, the replacement.
6. **Apply the GEO finding to two sections at least.** Aggarwal et al. measured up to 40% more visibility in generative engines from added statistics, quotations and cited sources. Every figure carries a source or becomes a `[NEEDS FIGURE]` marker.
7. **Refuse to guess** the publish date, author, any client result, price or statistic not supplied or publicly sourced.

## Then Codex does

Codex takes the bulk half: sweeping every site file for other pages on the same query, listing stale dates and figures, applying the fix list character for character, mapping inbound internal links, producing the diff. Bulk find and replace, rules already set.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Claude fills in first: {PAGE_URL}, {TARGET_QUERY}, {VERDICT}, {FIX_LIST}, {SITE_FOLDER}.

## Claude checks

1. **Read the diff, not Codex's summary.** Codex has been caught reporting no change while its edit sat in the tree.
2. Every new statistic carries a source. Unsourced number, reject.
3. No date invented. Swapping a year on a real event is fabrication, not a refresh.
4. The title link is unique on the site, no repeated boilerplate. Google calls `Foobar, foo bar, foobars` spam-like.
5. Redirects run one way, no chain.
6. Word count was not the goal; Google states length alone does not rank.
7. Reject and rerun naming the failing check. Never hand-patch and call it verified.

## Rules

- Public information only.
- Never invent a fact, a number or a quote. `[NEEDS FIGURE]` beats a plausible figure.
- Anything sent in someone's name says whose name it is.
- No keywords meta tag; Google Search does not use it.
- Do not chase heading counts or order; Google states neither matters.
- E-E-A-T is not a ranking factor, per Google. Do not sell it as one.
- Never mark up content not visible on the page.

## If Codex is not installed

Claude does the sweep, the rewrite and the diff itself, and says so in the output.

## Built from

- seranking/seo-skills, https://github.com/seranking/seo-skills, 135 stars on api.github.com, 5 Sep 2026; the four verdicts.
- inhouseseo/superseo-skills, https://github.com/inhouseseo/superseo-skills, 305 stars on api.github.com, 5 Sep 2026; anti-slop ruleset.
- Aggarwal et al., GEO, https://arxiv.org/abs/2311.09735, ACM SIGKDD 2024, peer-reviewed; the 40% figure.
- Google Search Central, https://developers.google.com/search/docs/appearance/title-link; Google's own rules.

# Update an old page and win the ranking back - Codex task

You are doing the bulk half of a content refresh. Another assistant has read the page, compared it with the current top three results, and set the verdict and the edits. Do not re-argue the verdict and do not invent edits of your own.

## You are given

- `{PAGE_URL}` the live URL of the page being refreshed.
- `{TARGET_QUERY}` the single search query it should rank for.
- `{VERDICT}` one of KEEP, REFRESH, CONSOLIDATE, KILL.
- `{FIX_LIST}` numbered edits, each naming a section, the current text and the replacement text.
- `{SITE_FOLDER}` the local folder holding the site source files.

## Produce

1. `cannibalisation.csv` every file under `{SITE_FOLDER}` whose title, H1 or first 200 words target `{TARGET_QUERY}` or a near-identical phrasing. Columns: `file_path,title,h1,match_type,word_count`. Include the page's own file. If only one row results, still write the file.
2. `stale-references.csv` every year, date, price, statistic and "as of" phrase in the page file. Columns: `line_number,current_text,type,proposed_replacement,needs_human`. Mark `needs_human` YES and leave the replacement empty for anything `{FIX_LIST}` does not settle. Never guess a replacement figure.
3. The `{FIX_LIST}` edits written into the page file, character for character.
4. `inbound-links.csv` every other file linking to `{PAGE_URL}`. Columns: `file_path,anchor_text,line_number`. Change none of them.
5. `refresh.diff` a unified diff of every file you touched.

If `{VERDICT}` is KEEP or KILL, produce items 1, 2 and 4 only, apply no edits, and say so at the top of your return.

## Rules

- Change only what `{FIX_LIST}` names, plus dates it settles with certainty.
- Never invent a statistic, date, price, quotation or client result. Write `[NEEDS FIGURE]`.
- No keywords meta tag.
- Do not add, remove or reorder headings to hit a count. Do not pad word count.
- Create no redirects, delete no files, merge no pages. Consolidation is a human decision taken after reading `cannibalisation.csv`.
- Public sources only. Nothing from a logged-in tool, paid dataset or private document.
- If a file will not parse, or a section named in `{FIX_LIST}` is missing, record it and carry on with the rest.

## Return

Print, in order:

1. The verdict given, and whether you applied edits.
2. Row counts for each of the four CSVs.
3. Every `needs_human=YES` row from `stale-references.csv`.
4. Each `{FIX_LIST}` item number marked APPLIED, NOT FOUND or SKIPPED, with a one-line reason for anything not applied.
5. Every `[NEEDS FIGURE]` marker inserted, with file and line number.
6. `git diff --stat` for the changed files, then the path to `refresh.diff`.

Print values read from the files. Do not retype or summarise them from memory.