AI Guides › Playbooks

The Plugin Build Card: Packaging One Recurring Client Job as a Claude Plugin

By Nigel Guy · 7 min read

Most small businesses already know AI exists. What they lack is the afternoon it takes to turn "we should use Claude for this" into something that runs the same way every week. The usual response from people selling AI services is to build something impressive and then look for a buyer, or to sell "AI consulting" so broadly that nobody can picture it. Both feel like progress; neither gets a yes.

The rule: package one recurring job that a named kind of client already does by hand, prove it with a demo on their own material, and charge for the setup rather than the idea.

What a Claude plugin actually is

A plugin is a folder bundling skills (instructions Claude follows when a task matches), commands, connectors and, in Cowork and Claude Code, agents. It needs a manifest at .claude-plugin/plugin.json; each skill is a SKILL.md in its own folder under skills/.

Fact Why it matters to you
Plugins work on paid plans: Pro, Max, Team and Enterprise Your client needs their own paid Claude account. Pro is roughly £18 a month in the UK at time of writing; the exact figure shows at checkout
You add one from Customize > Plugins, via Discover, Add > Add marketplace or Add > Upload plugin You can hand a client a .zip or .plugin file and walk them through Upload plugin
Add > Create with Claude builds a plugin from a conversation, with a Save plugin button at the end You can draft the first version without writing any files
Skills and commands load everywhere; agents and hooks only in Cowork and Claude Code Design for chat, the plainest surface
Uploaded plugins skip Anthropic's directory review Your client is trusting you, so keep it readable

The Plugin Build Card

One card per plugin, filled in before you build. A blank line means it isn't ready.

Line Question Weak answer Strong answer
Job What does someone do by hand, repeatedly? "Marketing" "Turns each weekly podcast episode into social posts"
Buyer Who exactly does this job and dislikes it? "Small businesses" "Two-person coaching firms with a weekly podcast"
Input What goes in? "Their content" "One episode transcript, pasted or attached"
Output What comes out, in what format? "Posts" "Ten ideas, each with a script, a LinkedIn post and a caption"
Left out What does it deliberately not do? Nothing "Doesn't schedule or publish; a person approves every post"
Demo What will you show them, on whose material? "A slide" "Last week's episode, run live"

Three jobs fit the card well. Each prompt below becomes the body of a plugin's SKILL.md.

Plugin 1: one long piece into short pieces, in the client's voice

The failure mode is generic output that could come from any marketing account.

You are a content editor working for [CLIENT_NAME], who publishes [CONTENT_TYPE: podcast / blog / video] for [AUDIENCE].

Source material: [PASTE_TRANSCRIPT_OR_ARTICLE]
Voice notes or past posts to match: [PASTE_2-3_EXAMPLES_OR_"NONE"]

Goal: ten short pieces the client could post after light editing.

Steps:
1. Read the full source before writing anything.
2. Choose the ten ideas or lines that stand alone without the rest of the source, strongest first.
3. For each, produce:
   a) A spoken script of 60–90 seconds whose opening line works as a hook.
   b) A LinkedIn post of 150–200 words.
   c) One caption line for Instagram or TikTok.
4. Reuse the speaker's own words and turns of phrase. Avoid stock marketing vocabulary.

Output: a numbered list. For each item show the source idea (with a timestamp or quote if available), then a), b) and c).

Rules:
- Do not add facts, figures or quotes that are not in the source.
- Mark any claim you cannot confirm from the source with [CHECK].
- If the source is too short for ten strong ideas, say so and ask how many to produce.

Before answering, check: ten items, three formats each, word counts respected, every [CHECK] flag in place.

Fill in: client name, content type, audience, the source, and any past posts that show their voice.

Plugin 2: research a real company, then draft a short first message

The failure mode is invented personalisation: a "specific" detail that was never actually found.

You are a research assistant helping [YOUR_NAME], who offers [YOUR_SERVICE], prepare a first message to one company.

Target: [COMPANY_NAME], website [COMPANY_URL]. Person to contact, if known: [NAME_AND_ROLE].

Steps:
1. Search the company's website and recent public news. Note what they sell, who to, and anything dated in the last three months.
2. Pick one detail that someone could only mention if they had actually looked, and that relates to [YOUR_SERVICE].
3. Draft a message of three to five sentences: open with that detail, name one problem you can help with, end with a single direct question.

Output:
- Research notes as bullets, each with its source URL and date.
- The chosen detail and why it was chosen.
- The draft message.

Rules:
- Every fact must have a link you actually opened. No link, no fact.
- If nothing specific and recent turns up, say so and stop; do not write a generic message dressed up as a personal one.

Before answering, check that each claim in the message appears in the research notes with a source.

Fill in: your name and service, the company, its website, and the contact if you know them.

Don't count on Claude reading LinkedIn, much of which sits behind a login; paste anything from there yourself.

Plugin 3: intake form answers into a draft proposal

The failure mode is internal notes posing as a proposal, or prices invented to fill gaps.

You are a proposal writer for [YOUR_BUSINESS], which provides [SERVICE].

Client intake answers: [PASTE_FORM_ANSWERS]
Your price list or rate rules: [PASTE_PRICING_RULES]
Earliest start date: [DATE]

Goal: a proposal written to the client, ready for them to accept.

Sections, in order:
1. Your situation: one paragraph restating their problem, using their wording where you can.
2. Scope: what is included, then one item deliberately not included and why it is a sensible later phase.
3. Timeline: two to four milestones with dates calculated from [DATE].
4. Investment: a one-off setup fee plus an optional monthly maintenance fee, taken only from [PASTE_PRICING_RULES].
5. Next step: one action and a reply-by date.

Rules:
- Never invent a price, discount or date that is not derived from the inputs.
- If the intake answers leave scope unclear, list your questions at the top instead of guessing.
- Use British English and £.

Before answering, check that every number traces back to the inputs.

Fill in: your business and service, the intake answers, your own pricing rules and a start date.

Building it

  1. Open Customize > Plugins, choose Add > Create with Claude, and describe the job using your Build Card.
  2. Give Claude the relevant prompt above as the skill body, and have it turn the placeholders into questions the skill asks.
  3. Select Save plugin, then test it in a fresh chat with real material.
  4. When it behaves, ask Claude in the same conversation for the plugin as a folder, check it contains .claude-plugin/plugin.json and skills/, and zip it for the client to add through Add > Upload plugin.

In Claude Code, claude plugin validate ./your-plugin checks the manifest.

A worked example (hypothetical)

Say you approach a two-person bookkeeping practice that writes three proposals a week by hand from a web form. Build Card: job, proposal drafting; input, form answers plus their rate card; output, a five-section proposal; left out, sending anything or touching their accounts software. You build Plugin 3, run it with permission on one of last month's real enquiries, and show the draft beside the proposal they actually sent. You quote a fixed setup fee (build, an hour of training, one round of changes) plus an optional monthly fee for keeping rates and prompt current. The figures are yours to set; this guide doesn't suggest any.

What to skip

Guardrails

Sources

All 751 AI guides · JulieMango plans from £17/mo