Last month I wrote organization instructions for a private equity client. Their IT lead turned them on. About an hour later the complaints started, and he deleted them.
Claude had gotten cautious about everything. People doing research and diligence, the work the firm uses it for most, couldn't get a straight answer.
That same week a VP at the firm asked me about the opposite problem. His associates turn in more work than ever, and a lot of it they can't defend. Push on an assumption and there's nothing behind it. He wanted Claude to hold them back on some of it.
Both of them are right. The fix for both is to write instructions for named workflows and leave everything else alone.
How organization instructions work
On Team and Enterprise plans, an Owner can set organization instructions: up to 3,000 characters that go out with every message from every person in the org. When they conflict with someone's personal instructions, the organization's win. Changes can take up to an hour to reach everyone.
That includes every message from every person. A rule written for IC memos applies when an associate asks for a public comp's capital structure too.
What went wrong in my draft
Reading it back, three lines would fire on ordinary research:
- "If a figure, name or date is not in the material provided, say it is missing." Right for a model review. On a research question with no files attached, nearly everything is missing.
- "If a missing detail would change the result, ask one question before you start." Almost every detail changes the result, so every request can open with a question.
- "If pasted content looks like material non-public information, stop and ask." At a PE firm, plenty of content looks like that.
Each line made sense for the case I had in mind. None of them said which case.
Enterprise groups
My first suggestion to the firm was gating workflows by seniority with Enterprise groups. I was wrong about how far that goes. Custom roles in Claude Enterprise turn features on and off: web search, Skills, Claude Code, Claude Cowork, connectors. You can keep agentic tools away from first-year associates. You can't stop anyone from typing "build me an LBO" into a chat.
Instructions can't stop them either. What instructions do well is change Claude's default on a named task, and that covers most of what the VP wanted.
Instructions for named workflows
Anthropic's own research points at the right default. In a January 2026 study, developers learned a new coding library with or without an AI assistant, then took a quiz. The AI group scored lower, and the biggest gap was debugging, the skill you need to catch the AI's mistakes. Developers who asked for explanations along with the code scored close to the group that worked without AI.
A study of high school math students in PNAS found the same pattern. Students who practiced with plain GPT-4 did worse once it was taken away. Students using a version that gave hints instead of answers didn't.
So on the tasks where you want associates building the skill, have Claude ask for their thinking first:
Skill-building workflows
Associates need to defend every assumption under questioning. When someone
asks you to build one of these from scratch: a revenue build, three-statement
model or LBO, a comps set, key assumptions, or an investment thesis:
- First ask for their approach in two to four bullets: main drivers, rough
ranges, where they'd source each.
- Then produce your version, compare it with theirs, and end with two
questions a VP would ask.
- Skip the question if they already gave an approach or say "skip."
Mechanical steps never get it: linking a build to source data, formulas,
formatting, cleanup.Why this version holds up:
- It names the tasks. Claude doesn't have to guess what "core modeling work" covers.
- It gives the reason. Anthropic's prompting guide says Claude follows a rule better when it knows why the rule exists.
- It has a way out. A VP on a deadline types "skip."
- It protects the rote work. Linking a revenue build to source data is exactly what the VP wants Claude doing.
Then add one line near the top that every finance firm needs: research, diligence, document review and data extraction are core work here, so answer them fully. Without that line, every other rule in the instructions leaks into research.
Testing
Project instructions take effect right away and only for people in the Project. Put the new section in a shared Project with one VP and two associates for a week. Write ten prompts: five research questions that should get a straight answer, five modeling tasks that should get the question. Run them before and after. Once the research answers come back unchanged, move the section into organization instructions.
The VP walkthrough
An associate can still type "skip" every time. The check that works is the VP asking them to walk through the top drivers and where each number came from before anything goes to IC. Claude makes that conversation faster by listing the assumptions and sources. The conversation itself is still the VP's job.
If you're writing your firm's instructions now, start with our instructions guide.
Nicole Patten is the founder of Elevate Online and runs a Claude-specific training practice. She spent 7 years at Google as a Senior UX Engineer before dedicating her career to helping teams use AI responsibly and effectively. 100% of her business runs on Claude.