on-page-fixer

---
name: on-page-fixer
description: Audits pages and returns the exact replacement title, meta description, H1 and internal links, scored. Use before publishing, or when a page ranks below where it should.
---

# Fix the title, headings and links that are costing you rankings

A scored audit of every page, plus replacement text ready to paste.

## Claude does

1. Asks for: the URL or its HTML, the one query it should win, the country and language, and whether it is local.
2. Runs four fixed phases: pull, compare to intent, score, write fixes, so two runs compare.
3. Scores 1 to 10 on seven dimensions: intent match, title, headings, depth, internal links, entity coverage, conversion clarity. Six or below is rewritten; seven or above is left alone.
4. Rewrites the title as one line unique to the page, plus the meta description and H1: no sitewide boilerplate, no keyword repetition. Google names "Foobar, foo bar, foobars" as the spam pattern.
5. Judges headings on reader need, not shape: Google says heading order does not matter and there is no ideal number, so "H3 before H2" is never flagged alone.
6. Writes one self-contained answer block of 134 to 167 words near the top, so an AI engine can quote the page alone.
7. Names 3 to 5 internal links pointing at the page, with anchor text, most important first.
8. Refuses to guess rankings, traffic, prices, hours or addresses, and marks them [NEEDS INPUT].
9. Ends with a verdict: KEEP, REFRESH, CONSOLIDATE or KILL, plus the reason.

## Then Codex does

Codex takes the bulk pull: every URL fetched, then title, meta description, canonical, headings in order, word count and every internal link with its anchor, plus text repeated across pages. Mechanical, and error-prone by hand past five pages.

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

Claude fills in first: {URL_LIST}, {TARGET_QUERIES} (URL plus its query), {SITE_DOMAIN}, {OUTPUT_FOLDER}.

## Claude checks

1. Rows equal URLs. A missing row is a failed fetch, not a clean page.
2. A blank title or H1 is refetched against the rendered page: raw HTML says nothing about text written by JavaScript, so "missing H1" is never reported from source alone.
3. No fix adds a keywords meta tag (Google Search does not use it) or cites E-E-A-T, which Google says is not a ranking factor.
4. Duplicates checked across the whole set, not a sample; every statement traces to a CSV cell.

Reject rule: if a check fails, rerun Codex on the failing URLs only; never pass a partial CSV as complete.

## Rules

- A grep is a lead; a fetched page is a finding.
- Never print a weighting the source does not print itself.
- Do not rewrite anything scoring seven or above.
- Public information only.
- Never invent a fact, a number or a quote.
- Anything sent in someone's name says whose name it is.

## If Codex is not installed

Claude does the extraction itself and says so.

## Built from

- AgriciDaniel/claude-seo, https://github.com/AgriciDaniel/claude-seo, 16,344 stars (api.github.com, 5 Sep 2026).
- inhouseseo/superseo-skills, https://github.com/inhouseseo/superseo-skills, 305 stars; four-phase order and scorecard.
- seranking/seo-skills, https://github.com/seranking/seo-skills, 135 stars; KEEP / REFRESH / CONSOLIDATE / KILL.
- Google Search Central, https://developers.google.com/search/docs/appearance/title-link; the authority.

# On-page fixer - Codex task

You are extracting on-page facts for an SEO audit. You do not write copy and you do not give advice.

## You are given

- {URL_LIST} - one URL per line. Fetch every one.
- {TARGET_QUERIES} - a URL followed by the single query that page should win. Carry the query through to the output untouched.
- {SITE_DOMAIN} - the site's own domain, used to separate internal links from external ones.
- {OUTPUT_FOLDER} - where the files below are written.

## Produce

1. `{OUTPUT_FOLDER}/onpage.csv`, one row per URL, with these columns in this order:
   `url, target_query, http_status, final_url, title, title_length, meta_description, meta_description_length, canonical, meta_robots, h1_all, heading_outline, word_count, internal_link_count, external_link_count, first_internal_anchor, has_keywords_meta, fetch_error`
   - `h1_all` - every H1 found, separated by ` | `. Leave empty only if there genuinely is none.
   - `heading_outline` - headings in document order as `H1:text > H2:text > H2:text`, truncated at 400 characters.
   - `has_keywords_meta` - yes or no.
   - `fetch_error` - the error text if the fetch failed, otherwise empty.
2. `{OUTPUT_FOLDER}/duplicates.csv` with columns `field, value, urls, count` listing every title and every meta description that appears on more than one URL in the set.
3. `{OUTPUT_FOLDER}/links.csv` with columns `source_url, anchor_text, target_url, position_index, is_internal` for every link in the body, `position_index` counting from 1 in document order.
4. `{OUTPUT_FOLDER}/evidence/` holding the saved HTML of each fetched page, named after the URL slug, so every row can be re-checked.

## Rules

- Fetch each URL once, follow redirects, record the final URL. Wait 1 second between requests.
- If a page renders its title or headings in JavaScript, say so in `fetch_error` as `js_rendered_suspected` rather than recording an empty title as fact.
- Never fill a cell with a guess. An unknown value is empty, and the reason goes in `fetch_error`.
- Do not repair, rewrite or normalise page text. Copy it exactly as served, including odd capitalisation.
- Do not skip a URL because it 404s or times out. Write the row with the status and the error.
- Public information only: no logins, no paywalled pages, no scraping anything behind authentication.
- Never invent a fact, a number or a quote.

## Return

Print, as plain text, in this order:

1. `URLS_GIVEN=<n>` and `ROWS_WRITTEN=<n>`. These must match.
2. `FETCH_ERRORS=<n>` followed by one line per failing URL and its error.
3. `DUPLICATE_TITLES=<n>` and `DUPLICATE_DESCRIPTIONS=<n>`.
4. `KEYWORDS_META_FOUND=<n>`.
5. `JS_RENDERED_SUSPECTED=<n>` followed by those URLs.
6. The four output file paths.

Print nothing else. No summary, no recommendations.