AI Guides › Workbench
By Nigel Guy · 7 min read
Most people ask Claude money questions as one big question: "Is this share a good buy?" or "Am I spending too much?" The answer is fluent, balanced and built on whatever the model happened to find. It feels like research, but it's research theatre: you can't tell which parts were checked and which were filled in.
The rule: split the question into five narrow jobs, give each one a fixed checklist as a skill, and make every skill say what it could not verify. A skill gives you a first read. It doesn't decide anything.
A skill is a folder holding a SKILL.md file: a short YAML header (name and description) followed by instructions. A one-off question gets a different method every time. A skill runs the same checks in the same order, so this week's result can be compared with last week's. And each of the five jobs below needs different inputs and fails in a different way, which one mega-prompt blurs together.
| Skill | What it checks | Cost at time of writing | Best for | The catch |
|---|---|---|---|---|
| Holding watch | Last seven days of news, rating changes and company announcements for shares you already own | Works on Claude Free, which includes skills, web search and code execution | A weekly "anything I need to know?" check | Web search misses paywalled news |
| Earnings review | One set of quarterly results against the company's own guidance and the prior period | As above. Anthropic also publishes a free /earnings skill (Apache 2.0) |
Results day | Only as good as the documents you give it |
| Comparables table | A company's valuation multiples next to a handful of peers | Anthropic's /comps skill is free; its data connectors are paid services |
Seeing whether a valuation looks unusual, and why | Peer choice drives the answer |
| Statement check | Internal consistency of an income statement, balance sheet and cash-flow statement | Your own skill, any plan | Spotting numbers that don't tie before you rely on them | It checks arithmetic and red flags. It is not an audit |
| Budget sort | Categorises a month of your own spending and compares it with a target | Your own skill, any plan | A monthly money review | You're sharing personal data. Strip it first |
Anthropic's pricing page lists Pro in US dollars. I couldn't confirm the UK £ price, so check it, including VAT, at checkout.
On claude.ai (web or desktop):
holding-watch, and put a SKILL.md inside it.holding-watch/SKILL.md rather than a loose file.Keep name to 64 characters and description to 200, the help centre's limits for uploads. Claude reads the description to decide when to use the skill, so say plainly what it does and when.
In Claude Code, put the folder at ~/.claude/skills/holding-watch/SKILL.md and run it with /holding-watch. Skills stored there don't carry over to Cowork or cloud sessions.
This is the source idea, rebuilt. When you run it, give it tickers and your reason for holding each.
---
name: holding-watch
description: Weekly check on shares I already own. Finds the last 7 days of news, analyst rating changes and company filings, and says whether any of it touches my reason for holding.
---
You are a careful research assistant helping a private investor review shares they own. You never recommend buying or selling.
Inputs needed: [TICKERS_AND_EXCHANGES] and, for each one, [MY_REASON_FOR_HOLDING] in one sentence. If either is missing, ask for it before searching. Do not guess the exchange.
Steps:
1. For each ticker, search for items dated within the last 7 days: news, analyst rating or target changes, and official company announcements or regulatory filings.
2. Prefer the company's own announcements over press coverage. Note the date and source of every item.
3. Sort each item: touches my reason for holding / material but unrelated / routine noise.
4. Never forecast the share price.
Output, per ticker:
- Verdict in one line: "Needs a look" or "Nothing beyond normal noise"
- Table: date | item | source | category
- One sentence on whether my reason for holding still stands
- "Could not verify": anything you could not trace to an original source
Before answering, check that every item has a date inside the window and a named source. Remove any that don't.
Fill in the company name and attach the results announcement plus the previous quarter's.
---
name: earnings-review
description: Reviews one set of company results against its own prior guidance and the previous period, using only documents I attach. Lists what changed and what to check next.
---
You are an analyst writing a first-pass note on [COMPANY] results for [PERIOD], for a reader who will check your work.
Use only the attached documents. If the prior period or the guidance is not attached, say so and ask for it rather than recalling figures from memory.
Steps:
1. Extract revenue, operating profit, earnings per share, operating cash flow and net debt for both periods, with page references.
2. Compare each with earlier company guidance; note one-off items and accounting changes.
Output: a table (metric | this period | comparison | change | page), then three bullets on what changed and three questions to check next. No buy, sell or hold language.
Self-check: every number has a page reference, and every percentage change matches the two figures it comes from.
In Claude Code, install Anthropic's version instead of writing your own:
claude plugin marketplace add anthropics/financial-services
claude plugin install financial-analysis@claude-for-financial-services
Then run /comps. The repository's own disclaimer is the right frame: nothing in it is investment advice. If you build your own, make it list the peers, say why each was picked, and date and source every figure.
Fill in the company and attach the three statements for the same period.
---
name: statement-check
description: Checks that a company's income statement, balance sheet and cash-flow statement agree with each other, and flags red flags for a human to investigate.
---
You are checking the attached financial statements for [COMPANY], period [PERIOD], for internal consistency. Use code to do the arithmetic, not mental maths.
Checks, in order:
1. Balance sheet balances (assets = liabilities + equity).
2. Closing cash on the balance sheet matches the cash-flow statement.
3. Profit after tax links to retained earnings, allowing for dividends shown.
4. Receivables or inventory growing much faster than revenue.
5. Operating cash flow persistently below reported profit.
Output: a table (check | pass / fail / cannot test | figures used | page), then up to three red flags written as questions for a human. If a statement is missing, mark dependent checks "cannot test" and ask for it.
Self-check: rerun any failed check once before reporting it.
Fill in your target budget and attach a month's transactions as CSV, with account numbers and sort codes removed.
---
name: budget-sort
description: Sorts one month of my own spending into categories and compares it with my target budget. Use when I upload a transactions CSV.
---
You are helping me review my own spending. You are not a financial adviser.
Inputs: [TRANSACTIONS_CSV] and [TARGET_BUDGET_BY_CATEGORY]. If no target is given, ask for one instead of inventing it.
Steps:
1. Use code to load the CSV and total by category. Put anything you cannot categorise in "Unsorted"; do not guess.
2. Compare each category total with its target.
3. List the five largest single transactions and any repeating payments.
Output: a table (category | spent | target | difference), the Unsorted list and recurring payments, then two observations. No product recommendations.
Self-check: category totals add up to the CSV total.
Start with the job you already do by hand. Own shares: holding watch. Household cash flow: budget sort. Add earnings, comps and statement checks only if you actually read company documents.