Write it down once, so it's right every time
The result is as good as what you put in the folder and what you write down once. One folder per job, a CLAUDE.md with your voice, scorecards from simple to measured, and PDFs turned into text before Claude works on them.
In every team we train, the first tools people build show the same thing. The result is as good as what you put in the folder and what you write down once.
A sample report does most of the work. Put all the documents for a job in one folder, add one past report as the example and ask Claude to copy the format and write in your style, and the first draft usually gets the style and the layout right. Then two things go wrong. Asking it to "check your work" still leaves small mistakes, because a general check isn't enough: the checks need writing down one by one. And when you ask it to add something, it often writes far more than you need. Both are fixed by writing things down.
One folder per job
The case folder changes every time. The starting context and your design system stay in one place and get linked in, so you only keep one copy of each.
One folder per job, one copy of everything else
case-folder/ one per jobCLAUDE.md read first, every chatsources/ PDF + its .md textoutput/ what Claude writes
starting-context/ linked inINSTRUCTIONS.md the standing briefVOICE.md how you writeSCORECARD.md the checksexamples/ a good past reportdesign-system/ your brand, linked
The case folder changes every time. Starting context and design system are linked in, never copied.
Don't worry about getting the layout perfect. Use it, change it and tidy it with Claude when it gets messy.
CLAUDE.md and your voice
Claude reads CLAUDE.md on its own before it does anything in that folder, in every new chat. Keep the name exactly, and copy it into each folder where you want the same rules.
Whatever you keep correcting by hand belongs in this file. If it writes too much, have it describe a short piece of your writing and turn that description into instructions for itself. It understands its own words best, so let it write them and you approve them.
Read the short report in examples/. Describe how it's written: sentence length, how much detail, how it opens and closes. Then write instructions for yourself that make your writing as short as this, and add them to CLAUDE.md for me to approve.
When you add anything to a report I've written, highlight it so I can see what's yours. Add this rule to CLAUDE.md.
When results start to get worse in a long chat, start a new one. It begins again from CLAUDE.md with a clean head. Every few months, give your old .md files back to Claude and ask it to make them shorter and easier to use. The tools change, so your files should too.
Scorecards, from simple to measured
A scorecard is the job description you'd give a new hire, with the KPIs you'd judge them on. It's as small or as big as the job needs.
Scorecards, from simple to measured
Simple
A list in SCORECARD.md
No invented facts. Every finding has its source.
Measured
Checks with a number
Every document in sources/ is cited at least once.
Assessment
A table it fills in first
Documents 1 to 5 in sequence. Document 4 is missing.
The more you want it to run on its own, the more exact the scorecard has to be.
A list is one to ten things a good draft has, and you check it yourself as you read. Checks with a number are for review work where a client will spot a gap: give it a threshold and a way to count. A table it fills in first (counting the documents, listing the key facts, flagging the holes) is what an agent runs on its own across several cases.
How hard you check depends on the report. On a long report, check the key issues and findings. On a short one, check every detail, because the client accepts fewer mistakes.
The best scorecards come from your own corrections. Give back the revised version and say why you changed it, the way you'd brief a subcontractor. You can dictate it.
My corrected version is in output/final.docx. Compare it with your draft. I changed it because <say why>. Update SCORECARD.md so the next draft passes these checks without me making the same changes.
The agent scorecard tool walks you through your first one.
PDFs into text
PDFs are messy underneath, so Claude has to guess at what's in them and can leave things out. They're also heavy, and copying them into every case folder adds up fast. Have Claude convert each one to a text (.md) file once, then check the text against the original.
Convert every PDF in sources/ to a .md file with the same name. Keep the tables as tables. Then compare each .md with its PDF and list anything missing: pages, tables, numbers, footnotes.
For work that has to be complete, like litigation, put the comparison in your scorecard so it runs every time.
Link it, don't copy it
A workspace in VS Code lets each case see your starting-context folder without copying it. Open the case folder, drag starting-context/ into the side panel and choose Add Folder to Workspace. Save the workspace when you close. Change a file in starting-context/ and every case that links it sees the change. The workspaces guide has a screenshot for every step, and the design system guide does the same for your brand.
Use redacted, dummy or public material until your AI policy is agreed.