Define good before you hand a job to AI
Before you hand a job to AI, you need to know what a good result looks like. Point Claude at the right file, write your brand down once, check a PDF, a workbook and a Word letter the way you'd check a new colleague's work, then turn what you checked into a scorecard.
Before you hand a job to AI, you need to know what a good result looks like. In session 2 of the Praxis programme we watch Claude make real files in VS Code (a PDF, a spreadsheet, a Word letter and a chart) and check each one the way you'd check a new colleague's work. You leave with a brand file Claude reads every time, and a scorecard for one job that repeats.
The working example is Fenwick & Rowe, a made-up advisory firm.
Point Claude at the file
"Summarise the Q3 accounts" makes Claude guess which file you mean. Name the file and it reads that one, and it can tell you which page each number came from. In VS Code there are three ways to do it:
- Type @ and the file name. Start typing @q3 in the Claude box and pick the file from the list. It works for folders too: @sources/
- Send the open file. With a file open, press Option K on a Mac or Alt K on Windows. Select some lines first and it sends just those.
- Copy the path. Right-click any file, choose Copy Relative Path and paste it into your message.
Then give the folder a rules file, CLAUDE.md. Claude reads it at the start of every conversation in that folder, so the rules apply to every job without you typing them again. Fenwick & Rowe's says:
Before you make any document, read brand/BRAND.md and follow it.
Every number says the file and page it came from.
Can't find it? Write [NOT IN SOURCE]. Not sure? Write [CHECK].
Save finished files in output/. Never write over a file in sources/.
Write like our past letters in brand/examples.
Your brand, written down once
A design system sounds grand. For us it's one file, brand/BRAND.md, that says what your documents look like and how you write: colours as hex codes, fonts, where the logo goes, page layout, how charts look and your voice. Claude builds it from documents you already have, and every PDF, sheet and chart after that comes out in your colours.
Look at the documents in @brand/examples/ and write brand/BRAND.md: our colours as hex codes, fonts, where the logo goes, page layout, how charts should look and how we write. Ask me about anything you can't tell from the files.
Three files, and what good looks like
Claude can read and make PDFs, Excel workbooks and Word documents. A file that opens isn't the same as a file you'd send, so here's what we check before it goes anywhere.
Read a PDF, make a PDF
Read @sources/q3-accounts.pdf. Give me revenue, margin and cash, each with its page number. Then make a one-page summary following @brand/BRAND.md.
Good looks like: every number has a page and the page says that number, anything missing says [NOT IN SOURCE], and it fits on one page in your colours. Watch for numbers rounded differently from the accounts, and text cut off at the bottom.
Build a workbook
Turn @sources/q3-transactions.csv into a workbook: a tab per month, a summary tab with formulas, headers in our brand green.
Good looks like: click a total and you see a formula like =SUM, the total matches the accounts and a Notes tab says where the data came from. Watch for totals typed in as plain numbers, and #REF! or #VALUE! anywhere.
Draft in your template
Draft the Q3 client letter in @templates/client-letter.docx using the summary. Keep our styles. Put anything you're unsure of in a comment.
Good looks like: headings use the template's styles, it reads like your past letters and doubts sit in comments, out of the text. Watch for "I hope this finds you well" and bold or font sizes typed in by hand.
To make Excel and Word files, Claude writes a short script and asks to run it. Read what it says it will do, then say yes. Open every file before you trust it.
The same chart, twice
Same data, same request, one with the brand file and a clear brief. The first try, "make a chart of the report hours", came back titled "Chart 1", in six colours, with an axis starting at 2 that made one bar look 24 times shorter than another when the real gap was about 3 times.
The same chart, twice
First try: “Chart 1”
- Axis starts at 2, so June looks 24 times shorter than January
- Six colours mean nothing
Report time down from 6.5 to 2.2 hours
- Grey before, green after, values on the bars
- Axis from zero, and a source line
Same data and request. The second one had the brand file and five rules. Made-up example.
The second follows five rules:
- The title says what happened: "Report time down from 6.5 to 2.2 hours".
- Colours mean something: grey for before, green for after.
- Values sit on the bars, so nobody reads numbers off an axis.
- The axis starts at zero.
- A source line says where it's from, so anyone can check it.
Make a bar chart of report hours by month from @output/run-log.xlsx. Follow @brand/BRAND.md. Title it with what the data shows, put the value on each bar, start the axis at zero and add a source line underneath. Save it as a PNG in output/.
From examples to a scorecard
Everything we checked becomes a scorecard: the standard every piece of AI work on that job gets judged against. Each check is something a colleague could tick without asking you. Here's the one for Fenwick & Rowe's monthly client summary:
A scorecard for the monthly client summary
No invented facts
Open 3 cited pages and find the number
All 3 match
Complete
Compare with last month's summary
Every section there, gaps marked
Totals add up
Click the totals in the workbook
Formulas, matching the accounts
On brand
Put it next to a past report
Same colours, fonts and logo
Sounds like us
Read the first paragraph aloud
You'd send it without rewording
Four out of five is a fail. Fix the brief or the brand file, then run it again.
"No invented facts" is always check number one.
I do this job every [week / month]: [describe it]. Here's a good example I made: @[file]. Help me write 5 checks for a good result. Each check needs a way to test it and a clear pass. "No invented facts" is check 1.
Build your own with the agent scorecard tool, and use redacted or dummy files until your AI policy is agreed.