web
Skill2 starsChanged 5 days agoMITView on GitHub
Audit and repair a static site's SEO/AEO assets — metadata, structured data, sitemap, robots, llms.txt, markdown mirrors. Checks they are present, correct, and still current against the pages they describe.
12 other projects use a version of this.
Adopted by
Most recently taken up by
- G4bar/web-clitheir own version, first seen Nov 15, 2026
- MysterionRise/ctf-kittheir own version, first seen Oct 13, 2026
- arch3rPro/ark-spacetheir own version, first seen Oct 11, 2026
- kjsa7412/claude_code_newslettertheir own version, first seen Oct 9, 2026
- jircdev/crew-plugintheir own version, first seen Oct 8, 2026
- Boxline-dev/examplestheir own version, first seen Oct 8, 2026
Other versions12 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 12
- MysterionRise/ctf-kit★ 12same name
- Agents4Academia-AI/UReKA★ 9same name
- LocalKinAI/kinclaw★ 9same name
- Shad0wMazt3r/The-Scaffolding★ 3same name
- arch3rPro/ark-space★ 1same name
- Boxline-dev/examples★ 1same name
- jircdev/crew-plugin★ 1same name
- kjsa7412/claude_code_newsletter★ 1same name
- toannhu96/korean-interview-coach-skill★ 1same name
- G4bar/web-cli★ 0same name
2 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository24 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- 1passwordSkill
- brainSkill
- brainstormSkill
- bridgeSkill
- claude-switcherSkill
- clear-cacheSkill
- commitSkill
- cross-artifactSkill
- ctoSkill
- diskSkill
- killSkill
- lookSkill
- parseSkill
- pbarSkill
- planSkill
- pruneSkill
- pushSkill
- reportSkill
- researchSkill
- swarmSkill
- vibe-auditSkill
- video-editorSkill
- wikiSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. 0 are common, 8 are distinctive, and 7 are very specific to this project.
0
Common
Most of the 11 edited versions we read keep these.
8
Distinctive
This author's own choices. Worth a look.
7
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 47
Add --live when the argument asks for it.
- Very specific: line 54
If the scanner reports it cannot find a site directory, ask which directory holds the built site and offer to record it in .claude/kit.json …
- Distinctive: line 60
Read every finding before you fix any of them.
- Very specific: line 69
Read both: a mirror that deliberately condenses a paragraph for machine readers is a decision to record as an ignore rule, while a mirror …
- Distinctive: line 76
Never fix a finding you have not looked at in its source file.
- Distinctive: line 136
Run the scanner again. Confirm the count dropped and that you introduced nothing new.
- Very specific: line 142
When cowork/brain/BRAIN.db exists, log a milestone summarizing the run, and log any finding the user declined to fix now as a task so it is …
- Very specific: line 229
Never silence a finding to make a run look clean.
- Distinctive: line 234
Scan before speaking: Never characterize the site's SEO/AEO state from reading a file or two.
- Distinctive: line 236
The scanner never writes: All repair goes through Claude's edits, so it is reviewable and matches each file's house style.
- Very specific: line 240
Manual means ask: Never rewrite a meta description, FAQ answer, or llms.txt paragraph and mention it afterwards.
- Distinctive: line 242
The page is the source of truth: When derived assets disagree with rendered copy, sync the asset to the page — unless the user says the …
- Distinctive: line 247
Report failures plainly: If findings remain, say which and why.
- Distinctive: line 249
Do not invent assets the project has not opted into: If a site has no markdown mirrors, their absence is not a finding — mirrors are only …
- Very specific: line 257
If .claude/kit.json has a rules."web" entry, read it and apply it as an additional instruction for this skill.
“Common” is measured: kept by at least 6 in 10 of the 11 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Site root
!ROOTS=$(jq -r '.web.roots[]?' .claude/kit.json 2>/dev/null); [ -z "$ROOTS" ] && ROOTS="web public dist site build _site out sitemd/site"; for r in $ROOTS; do [ -d "$r" ] && ls "$r"/*.html >/dev/null 2>&1 && echo "$r ($(find "$r" -name '*.html' -not -path '*/.*' | wc -l | tr -d ' ') html files)" && break; done || echo "No site directory found — set web.roots in .claude/kit.json"
Base URL
!jq -r '.web.baseUrl // "not set — set web.baseUrl in .claude/kit.json for canonical and sitemap checks"' .claude/kit.json 2>/dev/null || echo "no kit.json"
Asset inventory
!ROOT=$(jq -r '.web.roots[0]? // empty' .claude/kit.json 2>/dev/null); [ -z "$ROOT" ] && for r in web public dist site build _site out sitemd/site; do [ -d "$r" ] && ROOT="$r" && break; done; for a in robots.txt sitemap.xml llms.txt llms-full.txt _headers; do [ -f "$ROOT/$a" ] && echo " present $a" || echo " MISSING $a"; done 2>/dev/null
Web — SEO/AEO asset audit
Static sites accumulate two kinds of rot. Assets go missing (no llms.txt, no structured data, no snippet directives), and — more insidiously — assets go stale: the sitemap still claims a page was last touched in March, the markdown mirror describes a service you renamed, the FAQ schema answers a question the page no longer asks. The second kind is worse than having nothing, because it feeds confident wrong answers to the engines reading it.
This skill scans for both, fixes what has one right answer, and asks about anything that involves words a human would recognize as their own voice.
Dispatch
| Argument | Behavior |
| (empty) | Scan, auto-fix mechanical findings, propose the rest |
report | Scan and report only — write nothing |
live | Scan plus probe the deployed site, then fix as normal |
live report | Probe and report, write nothing |
Procedure
Step 1 — Scan
python3 .claude/skills/web/scan.py --jsonAdd --live when the argument asks for it. The scanner reads web.roots and
web.baseUrl from .claude/kit.json; pass --root / --base-url to override.
It writes nothing, ever. Output is JSON: {root, baseUrl, pages, counts, findings},
each finding carrying severity, check, file, line, message, detail,
and fix (auto or manual).
If the scanner reports it cannot find a site directory, ask which directory holds
the built site and offer to record it in .claude/kit.json rather than passing
--root every run.
Step 2 — Triage before touching anything
Read every finding before you fix any of them. The scanner is deliberate about severity but it cannot see intent:
- A
noindexpage with no Open Graph tags is correct, not a gap. The scanner already suppresses this — if you see it flagged, the page's robots meta is not what you assumed. mirror-missing-sectionon a heading the mirror deliberately rewords is a judgment call, not a defect.mirror-text-driftreports the page's wording and the mirror's side by side. Read both: a mirror that deliberately condenses a paragraph for machine readers is a decision to record as anignorerule, while a mirror still carrying copy the page has since rewritten is exactly the rot this check exists to find.schema-price-mismatchmay mean the page copy changed and the schema is now lying, or that the page says "starting at $5,000" while the schema correctly usesminPrice. Read both before deciding which one is wrong.
Never fix a finding you have not looked at in its source file.
Step 3 — Auto-fix the mechanical findings
Findings marked "fix": "auto" have exactly one correct answer. Apply them
directly, no confirmation:
| Check | Fix |
sitemap-stale-lastmod | Set <lastmod> to the page's real last-modified date |
sitemap-missing-page | Add the <url> entry |
sitemap-orphan | Remove the entry, or note it if the page is deploy-generated |
sitemap-no-lastmod | Add <lastmod> |
open-graph | Add the missing tag, mirroring existing values on the page |
og-image-relative | Make the URL absolute against baseUrl |
twitter-card | Add summary_large_image when an og:image exists, else summary |
snippet-directives | Add max-snippet:-1, max-image-preview:large, max-video-preview:-1 |
canonical, canonical-relative, canonical-mismatch | Point at the absolute page URL |
headers-content-type | Add a Content-Type rule for the asset |
robots-no-sitemap, robots-no-llms | Add the reference |
favicon, apple-touch-icon, viewport, lang | Add, reusing assets already in the project |
llms-full-stale | Regenerate from the current mirrors |
Match the file's existing formatting — indentation, attribute order, quote style, whether tags are one-per-line or wrapped. A fix that reformats surrounding lines is a bad fix.
Step 4 — Propose everything else
Findings marked "fix": "manual" touch words, structure, or claims. Show the
user what is wrong and what you would write, then wait.
This especially covers anything a human would recognize as their own voice —
meta descriptions, llms.txt prose, FAQ answers, page copy. Many projects hold
the rule that the human writes all customer-facing language; machine-facing text
like llms.txt still counts, because it is the text an answer engine is most
likely to paraphrase when describing them.
Present them grouped, with the actual before/after:
NEEDS YOUR CALL
faq-answer-drift — web/index.html
Page: "Nope. For right now, it's just me."
Schema: "Nope. For right now it's just me."
→ The schema is stale. Update it to match the page? [y/n]
description-length — web/index.html (179 chars, aim ≤165)
Current: "I handle all of your tech so you can focus..."
Proposed: "..."
→ Your copy. Rewrite, or leave it? [y/n/edit]For content drift, the page is almost always the source of truth and the derived
asset (schema, mirror, llms-full.txt) is what is stale. Say which direction you
propose to sync and why, because occasionally it is the other way round.
Step 5 — Re-scan
Run the scanner again. Confirm the count dropped and that you introduced nothing new. Report the before/after counts honestly — if something is still failing, say so rather than describing the run as clean.
Step 6 — Record, if the project has a brain
When cowork/brain/BRAIN.db exists, log a milestone summarizing the run, and log
any finding the user declined to fix now as a task so it is not silently dropped:
INSERT INTO logs (type, title, body, tags, importance)
VALUES ('milestone', 'Web audit: <n> fixed, <n> deferred', '<what changed>', 'aeo,seo,website', 5);Normalize tags: split on comma, trim, lowercase, sort, dedupe, rejoin.
What gets checked
Per page — lang, title presence and length, meta description presence and
length, viewport, canonical (present, absolute, matches the page URL), robots
snippet directives, the Open Graph set, Twitter card, favicon and
apple-touch-icon, exactly one h1, heading-level skips, img alt text, JSON-LD
parses, and JSON-LD @id references resolving to a node that exists.
Site-wide — robots.txt exists and does not block the site, points at a
sitemap, mentions llms.txt when one exists; sitemap.xml is valid XML, lists
every indexable page, lists no page that does not exist, lists no noindex page,
and carries lastmod on each entry; llms.txt follows the llmstxt.org shape and
every link in it resolves on disk; _headers declares a Content-Type for
.md assets when nosniff is set.
Currency — this is the part that catches rot rather than absence:
sitemap-stale-lastmod—<lastmod>older than the page's real last change, taken from git when the file is committed and clean, filesystem mtime otherwisemirror-missing-text/mirror-text-drift— a paragraph of page copy with no counterpart in the markdown mirror, or one the mirror words differently. This is the check that catches an edited sentence: mirrors are compared block by block, because a whole-document similarity score cannot see one rewritten line in a thousand-word page and reports clean while the mirror still tells engines the old storymirror-drift/mirror-missing-section— whole-document and heading-level backstops for the same mirrors, catching wholesale rot and dropped sectionsllms-full-stale— the concatenated corpus missing any substantive paragraph of a current mirrorfaq-count-mismatch/faq-question-mismatch/faq-answer-drift—FAQPagestructured data disagreeing with the<details>blocks actually renderedschema-price-mismatch— a price asserted in structured data that appears nowhere in the visible copyllms-txt-dead-link—llms.txtpointing at a file that does not exist
Live (--live only) — served Content-Type for .md and .txt assets,
canonical Link headers on markdown mirrors, HTTP status for every page and
asset, and redirect chains. This is the only way to catch deploy-config drift,
where the local _headers looks right and production ignores it.
Configuration
{
"web": {
"roots": ["web"],
"baseUrl": "https://example.com",
"ignore": [
{
"check": "mirror-missing-section",
"file": "web/index.md",
"match": "Hi, I’m Tyler",
"note": "the mirror deliberately uses a descriptive heading for machine readers"
}
]
}
}roots is a list of candidate directories; the first that exists and contains
HTML wins. Absent, the scanner tries web, public, dist, site, build,
_site, out, sitemd/site, then the project root.
baseUrl is the canonical origin. Without it, canonical-URL and sitemap-URL
checks are skipped and --live cannot run — everything else still works.
ignore records divergences the project has decided are correct, so a re-run
does not re-litigate them. Each rule needs check; file and match (a
substring of the finding's message) narrow it; note explains the decision and
is printed back whenever the rule fires.
Suppressed findings are counted in every run's header and listed under
--show-suppressed, and they stay in the JSON under suppressed. They are never
silently dropped — a suppression you cannot see is indistinguishable from a check
that stopped working.
Only add an ignore rule when the user has decided the current state is right.
Never silence a finding to make a run look clean. A rule with no note is a
rule nobody will be able to evaluate in six months.
Rules
- Scan before speaking. Never characterize the site's SEO/AEO state from reading a file or two. Run the scanner.
- The scanner never writes. All repair goes through Claude's edits, so it is reviewable and matches each file's house style.
- Auto means auto. Findings marked
autoare applied without asking. Do not pad the conversation by confirming a missingog:site_name. - Manual means ask. Never rewrite a meta description, FAQ answer, or
llms.txtparagraph and mention it afterwards. Propose, then wait. - The page is the source of truth. When derived assets disagree with rendered copy, sync the asset to the page — unless the user says the page is the thing that is wrong.
- Verify by re-scanning. A fix is not done because the edit applied. Re-run and confirm the finding is gone.
- Report failures plainly. If findings remain, say which and why. Never describe a run as clean when it is not.
- Do not invent assets the project has not opted into. If a site has no markdown mirrors, their absence is not a finding — mirrors are only checked when a page declares one.
Project overrides
If .claude/kit.json has a rules."web" entry, read it and apply it as an additional
instruction for this skill. Absent file or key means no overrides — that is the normal case.
jq -r '.rules."web" // empty' .claude/kit.json 2>/dev/nullDoes it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

