The five Claude skills I actually use

page-audit

---
name: page-audit
description: Score a page against the live top-10 SERP for a target keyword and return a deterministic 0-100 optimisation score with a ranked gap list. Use when asked to "audit this page", "score this page for [keyword]", "what do I need to rank for [keyword]", or when a page is underperforming and nobody knows why.
---

# On-Page Audit

Answer one question: what is on the pages that outrank us, that is not on ours?

## Inputs
- Target URL
- Target keyword
- Market (country + language) — the SERP is localised, and auditing
  against the wrong market is worse than not auditing at all

## Steps
1. Pull the top 10 organic results for the keyword in the target market.
   Exclude the target URL if it already ranks.
2. Fetch and parse each competitor page: title, H1, H2s, word count, and
   term frequency for the head term and its variants.
3. Parse the target page the same way, with the same parser. Different
   parsers produce different numbers and destroy the comparison.
4. For each measured term compute the range across competitors
   (min / median / max) — never a single target number. Ranges are honest.
   Single numbers pretend to a precision that does not exist.
5. Score the target 0-100 in code against those ranges. Suggested weights:
   coverage of terms all top-3 use > body length inside range >
   heading placement > everything else.
6. Emit the gap list ranked by (weight x distance from median).

## Output
| Field      | Value                                                  |
|------------|--------------------------------------------------------|
| Score      | 0-100                                                  |
| Band       | Critical <40 / Weak 40-59 / Competitive 60-79 / Strong 80+ |
| Top gaps   | 5 max, each with current value vs competitor range     |
| Word count | target vs competitor median                            |

## Rules
- The score is deterministic code. Never adjust a score by judgement, and
  never present an estimate as a score. The moment a score becomes a vibe
  it stops being evidence you can put in front of a client.
- Store every run with its date. The second audit is worth more than the
  first, because it is the one that proves movement.
- Recommend nothing you cannot point at in a competitor's HTML.
- Always report the date and the SERP the score was measured against.
  SERPs move; a score without a date is a rumour.

gbp-audit

---
name: gbp-audit
description: Audit a Google Business Profile and its map-pack competitors, returning a deterministic 0-100 score across six pillars plus a ranked fix list with evidence tiers. Use for "audit this GBP", "why aren't we in the map pack", "check this business on Google Maps", or any local-SEO review.
---

# Local Profile Audit

## Read-only, always
This skill never writes to a live profile. Everything it does is
observation. A tool that can edit a client's public listing is one bad
tool call away from a very bad day.

## The six pillars
| Pillar      | What it measures                                        |
|-------------|---------------------------------------------------------|
| Foundations | Name, address, phone, hours, categories, service area   |
| Reviews     | Volume, velocity, rating, response rate and quality     |
| Risk        | Guideline violations that could suspend the listing     |
| Website     | Landing-page relevance and speed for the mapped query   |
| Conversion  | Photos, products, booking path, messaging enabled       |
| Offsite     | Citation consistency, obvious NAP conflicts             |

## Steps
1. Capture the profile as it renders publicly. Snapshot it — you need
   the before, and you will not be able to reconstruct it later.
2. Score each pillar 0-100 in code from the captured fields, then a
   weighted composite.
3. Tag every finding with an evidence tier:
   - Confirmed  — visible in the capture
   - Inferred   — consistent with the data, not directly visible
   - Hypothesis — worth testing, no evidence yet
4. Tag every recommendation with policy risk: Safe / Grey / Violates.
5. Capture the top 3 map-pack competitors for the primary query. Flag
   visible violations (keyword-stuffed names, virtual offices, duplicate
   listings) as redressal candidates.

## Output
Composite score, six pillar scores, a findings table
(finding / pillar / evidence tier / policy risk / fix), competitor flags.

## Rules
- Evidence tier on every single finding. An untiered finding is an
  opinion wearing a lab coat.
- Never emit a recommendation in the Violates band, even when it works.
  Short-term win, long-term suspension, and it is your name on the invoice.
- Snapshot before and after. Local work is slow, and proof is the only
  thing that survives the wait.

content-strategy

---
name: content-strategy
description: Plan SEO content that builds topical authority without cannibalising existing pages. Use for "what should we write", "content plan for [site]", "topical authority map", "find niche search terms", or before commissioning any piece of content.
---

# Content Strategy

The job is not "more posts". The job is to own a topic so completely that
the next piece has somewhere to link from.

## Step 1 — Inventory before ideas
List every existing URL and the query each one is actually trying to win.
You cannot plan additions until you know what you already have. Most
"content problems" are three thin pages fighting each other.

## Step 2 — One page, one intent
For every candidate topic, name the single query it targets and the intent
behind it (informational / commercial / transactional).
**If an existing page already targets that intent, you improve that page.
You do not write a new one.** Cannibalisation is self-inflicted and it is
the most common reason a site plateaus.

## Step 3 — Cornerstone and spokes
- Cornerstone: the definitive piece on the head term. Few of these. They
  earn links.
- Spokes: narrower questions that each link UP to the cornerstone.
Links point spoke -> cornerstone, never the reverse. Directional linking is
what turns a pile of posts into an authority structure.

## Step 4 — Earn the right to the claim
Every piece needs at least one thing competitors cannot copy: original
data, a real result, a screenshot of the thing working. A post assembled
entirely from other posts ranks like one.

## Output
| Field | Value |
|---|---|
| Topic | the piece |
| Target query + intent | one each |
| Type | cornerstone / spoke |
| Links to | the cornerstone it supports |
| Proof asset | the non-copyable thing it carries |
| Cannibalisation check | existing URL it might conflict with, or "clear" |

## Rules
- Never plan a page that competes with a page you already own.
- Cap cornerstones. A site with forty cornerstones has none.
- If you cannot name the proof asset, the piece is not ready to commission.
- Volume is a vanity input. Coverage of one topic beats presence in ten.

humanizer

---
name: humanizer
description: Remove the signs of AI-generated writing from text so it reads as written by a person. Use when editing or reviewing any draft, before anything client-facing or published, or when asked to "make this sound human", "clean this up", "does this read as AI".
---

# Humanizer

Based on the patterns catalogued in Wikipedia's "Signs of AI writing".
Run it on anything before it leaves the building.

## The tells, and the fix

**Inflated symbolism** — "stands as a testament to", "plays a vital role in",
"underscores the importance of". Cut the sentence or state the fact plainly.

**Promotional voice in neutral writing** — "rich cultural heritage",
"breathtaking", "seamless". If it reads like a brochure, it is one.

**Superficial -ing analysis** — "..., highlighting the need for",
"..., showcasing its versatility". These clauses append a conclusion the
evidence did not earn. Delete them; the reader draws their own.

**Vague attribution** — "industry experts say", "studies have shown",
"it is widely regarded". Name the source or drop the claim.

**Rule of three, everywhere** — "fast, reliable, and scalable". One real
adjective beats three decorative ones. Vary the rhythm deliberately.

**Negative parallelism** — "It's not just X, it's Y." Once is a device.
Twice is a tic. Three times is a signature.

**Em-dash overuse** — the single loudest tell. Most can become a full stop.

**Filler openers** — "In today's fast-paced world", "It's important to note
that", "When it comes to". These carry no information. Start at the point.

**Section-ending summaries** — a paragraph restating the paragraph above it.

**Hedging stacks** — "may potentially help to somewhat improve".

## Method
1. Read once for meaning. Note what the piece is actually claiming.
2. Pass for the tells above, one at a time. Fixing them all at once
   produces mush.
3. Read aloud. Anything you would not say out loud, rewrite.
4. Check the rhythm: if every sentence is the same length, vary them.

## Rules
- Never change a fact while fixing the prose. Style edits only.
- Preserve the author's actual voice — this removes machine tics, it does
  not impose a house style.
- Shorter is usually the fix. Most AI text is 20% padding by volume.

client-email

---
name: client-email
description: Write client-facing emails that land an uncomfortable number, deliver audit findings, or open a scope conversation. Use for "write this email to [client]", "how do I tell them about this", "turn these findings into something sendable".
---

# Client Email

## Structure
1. The point, in the first sentence. No "hope you're well" — they scan.
2. The evidence. One or two specifics with numbers. Not five.
3. What it means for them, in their language rather than yours.
4. The ask. Exactly one, with a date on it.

## Rules
- Name the bad number in the subject line or the first sentence. Burying
  it reads as hiding it, and they always find it eventually.
- Never blame a previous agency by name. Describe what was done and what
  it cost. The reader draws the conclusion, and you keep your reputation.
- One ask per email. Two asks reliably produces zero.
- No hedging stacks. "I think we might possibly want to consider" is four
  hedges saying nothing. Say the thing.
- If you cannot show it, do not claim it. Every number in a client email
  should be one you can produce on request inside a minute.

## Before sending
- Read it as the client having a bad day. Does any sentence read as
  blame-shifting?
- Cut every sentence that exists to make you feel better rather than to
  inform them.
- Check the ask is actually answerable in a single reply.