AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
The usual mistake is to start with the impressive version: the dashboard that pulls in every account automatically, the assistant that "runs your week", the seven-agent team with a name for each seat. It feels like progress because the screenshots look good. What you actually get is something you stop opening after a fortnight, because it was built around the demo and not around one job you already do.
The rule: build the smallest version that does one real job, use it for two weeks, and only then add the next piece.
Below are three starter builds, in order: an income dashboard from one prompt, one scheduled task, and a three-agent team in Claude Code.
| You need | For which build | Cost at time of writing |
|---|---|---|
| A Claude account with artifacts switched on | Income dashboard | Free plan works |
| A paid Claude plan (Pro, Max, Team or Enterprise) | Scheduled task | Pro is listed at US$20 a month; check the pricing page for the current UK price in £ before you subscribe |
| Claude Code (included with Pro and Max) and a folder on your computer | Agent team | Included in the paid plan above |
| A CSV export of last year's income, or a quick spreadsheet you type up | Income dashboard | Free |
Artifacts are the interactive pages Claude builds beside the chat. Go to Settings > Capabilities and make sure the code execution and file creation setting is on (the help centre currently calls it "Cloud code execution and file creation") (on Team or Enterprise plans an admin controls this under Organization settings > Capabilities). Artifacts can save data between visits, up to 20 MB of text per artifact, which is plenty for a few years of income lines.
Gather your income streams first: salary, client work, brand deals, digital product sales, anything else. A rough CSV with date, source, stream and amount is enough. Then paste this into a new chat.
You are a careful personal-finance tool builder. I am a UK-based [EMPLOYEE / SOLE TRADER / LIMITED COMPANY DIRECTOR] with several income streams.
Goal: build me an interactive artifact that keeps all my income in one place.
My income streams are: [LIST_OF_STREAMS, e.g. salary, client work, brand deals, digital products, other sales].
My currency is GBP. My year runs [TAX_YEAR_6_APRIL_TO_5_APRIL / CALENDAR_YEAR].
Sample data (or "none yet"): [PASTE_CSV_OR_TYPE_NONE]
Steps:
1. Confirm the streams and the year setting back to me in two lines.
2. Build a form to add an income entry: date, stream, payer, amount (gross), note.
3. Save entries so they persist between visits.
4. Show: total income for the selected year, a month-by-month breakdown, and streams ranked by total, each with its share of the whole.
5. Add a CSV export button so I always have my own copy.
Constraints:
- Do not estimate tax, National Insurance or profit; show gross income only and label it as such.
- Do not invent sample figures. If I gave no data, leave the dashboard empty.
- If any input above is missing or unclear, ask me before building.
Before you finish, check that every total matches the sum of the entries and that the export includes every field.
Fill in your status, your list of streams, which year you use, and either paste data or write "none yet".
If you file Self Assessment, choose the UK tax year so the view matches the period you report on.
Before you open the scheduler, write the task on paper in one sentence that starts with a verb and names its output: "Every Monday, summarise last week's unread client emails into a list of replies I owe." If you cannot name the output, the task is not ready to hand over.
Good first tasks are read-and-summarise jobs. Tasks that send, pay, delete or publish things are not first tasks.
Scheduled tasks are on all paid plans. In the Claude desktop app:
On the newer layout some Pro and Max accounts have (no separate Chat and Cowork options), describe the task and its frequency in a normal message, answer Claude's questions, review the proposal and click Schedule.
Use this as the task prompt:
You are my weekly admin assistant. Each run, do one job only: [THE_ONE_TASK, e.g. review last week's emails from clients].
Sources you may read: [CONNECTOR_OR_FILES, e.g. my Gmail label "Clients"].
Time window: [e.g. the previous seven days].
Output, in this order:
1. A list of items that need a reply from me, each with sender, subject, date and one line on what they want.
2. Anything that looks urgent or overdue, flagged at the top.
3. A one-line count of what you checked.
Rules:
- Read only. Do not send, delete, archive or label anything.
- If a source is unavailable, say so plainly instead of producing a partial list that looks complete.
- If my instructions are ambiguous, list your question at the top of the output rather than guessing.
Before finishing, check that every item you list came from the stated source and time window.
Fill in the single task, the source Claude may read, and the time window.
Scheduled tasks run remotely, so they keep going when your computer is asleep. They use your connectors and files saved to your Claude account, not a folder on your computer; a task that needs local files or apps runs only locally.
A "team" in Claude Code is a set of subagents: Markdown files with a short settings header that Claude delegates work to. Start with three seats, not seven. A useful first trio for content work is a researcher, a drafter and a checker.
In your project folder, create .claude/agents/ (or use ~/.claude/agents/ to make them available in every project). Add one file per agent. Here is the checker, which is the one most people skip:
---
name: checker
description: Reviews drafts for unsupported claims, wrong figures and missing sources. Use after any draft is written.
tools: Read, Grep, Glob
---
You are a fact and claims checker for [YOUR_BUSINESS_OR_TOPIC].
For the draft you are given:
1. List every factual claim, figure, price and product name.
2. Mark each one as SOURCED (with the source), UNSOURCED, or CONTRADICTED.
3. List any sentence that promises an outcome the draft cannot support.
Output a table with columns: claim, status, source or note.
Do not rewrite the draft. If you cannot find the draft or its sources, say so and stop.
Before replying, confirm every row has a status.
Fill in your business or topic. Write the researcher and drafter files the same way: one job each, a clear description of when to use it, and only the tools that job needs.
The name and description fields are required; tools is optional, and if you leave it out the agent inherits every tool. Giving the checker read-only tools keeps it from editing the work it is meant to judge.
The /agents command now just reminds you to edit these folders or ask Claude (versions 2.1.197 and earlier opened a wizard), so editing the files directly is the reliable route.
@"checker (agent)" review draft.md, and confirm the reply is the claims table and not a rewrite.Agent tool out of its tool list.