Free Guide · No setup required

Where do Claude instructions go?

Claude reads instructions from four places - your organization, you, one Project, and one Skill. Most people put the wrong things in the wrong place. This guide covers what belongs where, with examples you can copy and a test set to run before everyone gets them.

Written for individuals setting up their own Claude account and for teams rolling out Claude Team or Enterprise. The four scopes are the same for both. What changes is who a rule reaches.

Last updated November 8, 2026

Open the generatorsSkip the reading - build yours in 3 minutes.

How they combine

Precedence

Claude reads every scope that applies, all at once. Anthropic documents one rule for conflicts: when organization and personal instructions directly contradict each other, the organization wins. Anthropic adds that in rare edge cases, behavior may vary.

There is no published order for Project or Skill instructions. So don't rely on one. Write each scope so it doesn't contradict the ones above it, and test any rule that matters before everyone gets it.

Scope · Everyone in the org

Organization Instructions (Team and Enterprise only)

Applies to: Everyone in your Team or Enterprise organization, on every message they send. Only Owners can see the text, so nobody else can tell why Claude behaves the way it does.

Where you set it: Team or Enterprise. An Owner or Primary Owner sets them in Organization settings → Organization and access. Up to 3,000 characters, and changes can take up to an hour to reach everyone.

Put here

  • Your company name, what you do, who you serve
  • The work Claude should always answer in full: research, diligence, document review
  • A sources-and-assumptions norm for anything that goes into a deliverable
  • What Claude does when sensitive material shows up: say so once, then keep working
  • Brand voice rules (sentence shape, banned phrases, tone)
  • Common references everyone needs (the product taxonomy, the customer segments, the OKR framework)

Don't put here

  • “Never paste” rules for people. Claude reads this file on every message, so those lines turn into warnings whenever someone shares a working file. They belong in your AI use policy.
  • “Stop and ask” rules meant for one workflow. Scope them to a Project and name the tasks.
  • Anything specific to one person's role (that's Instructions for Claude)
  • Anything specific to one project (that's project instructions)
  • A model preference. Instructions can't set the model; the org default is its own admin setting.

Example you can adapt

# COMPANY CONTEXT
We are [Company Name], a [middle-market private equity firm] investing in [sectors]. Most people here do research, diligence, modeling and portfolio work.

# CORE WORK
Research, diligence, market sizing, comps, document review and data extraction are core work here. Answer them fully and directly. When information is incomplete, give the best answer from what is available, state your assumptions, and keep going.

# SOURCES AND ASSUMPTIONS
Our work goes to investment committee and investors, so reviewers need to check it fast. For any figure, name or date that will go into a deliverable, say where it came from: the file and section, a cited source, or your own estimate.

# SENSITIVE MATERIAL
People here share confidential deal and portfolio material as part of normal work. Work with it normally. If a request looks like it relies on material non-public information about a public company, say so in one sentence, suggest checking with Compliance, and continue with the parts that do not depend on it.

# VOICE
Direct. Plain English. Short sentences. American English.

Why the org example is written this way →

Scope · You, everywhere

Instructions for Claude (your global instructions)

Applies to: Only you. Every conversation in your account, including every Project.

Where you set it: Every plan, including Free. In Settings → Your Claude (Account in older layouts), under Instructions for Claude.

Put here

  • Your role, function, and what you're trying to do
  • How you want Claude to talk to you (push back, don't validate; cite sources; be direct)
  • Format preferences (no em-dashes, no bullet lists unless asked, etc.)
  • Stable facts about you and your work that come up often (your name, title, region, key relationships)
  • Your stress signals or context (e.g. 'I prefer brevity when I'm in back-to-back meetings')

Don't put here

  • Anything project-specific - those instructions live on the Project itself
  • Anything that changes often (project status, current priorities for this quarter - put those in your Strategic Advisor Project as a knowledge file you update)
  • Confidential org context you wouldn't want surfacing in unrelated conversations

Example you can adapt

# ABOUT ME
I'm [Name], [Role] at [Company]. I work in [time zone], I'm time-poor, and I think in 5/20/30-minute pockets.

# HOW TO TALK TO ME
Push back instead of validating. If you don't know something, say so and ask the question that would change your answer. Be direct. Use plain English. Sentences are short.

# FORMAT
No em-dashes. No bullet lists unless I ask for them. When I ask "what should I do," give me one recommendation and the main tradeoff.

# CITATIONS
When pulling from my uploaded files, name the source ("from your Q3 priorities doc"). When you don't have a source, say so.

Scope · One Project

Project Instructions

Applies to: Anyone working inside that Project, and nowhere else. Changes take effect right away. That makes a Project the right place to test a rule before it goes org-wide.

Where you set it: Every plan. Free accounts can create up to five Projects. Sharing a Project with colleagues needs Team or Enterprise.

Put here

  • The role this Project plays - 'Strategic Advisor,' 'M&A triager,' 'Voice profile for external comms'
  • The specific domain context this Project covers (and ONLY this Project)
  • How this Project should behave differently from a default Claude conversation (e.g. more skeptical, more concise, more research-heavy)
  • Workflow rules that should fire only on named work, like asking an analyst for their approach before Claude builds a model
  • Format rules unique to this Project's outputs (e.g. memos always in a 6-section decision format)

Don't put here

  • Things that should apply to every Project - those go in Instructions for Claude
  • Long reference content - that goes in knowledge files inside the Project, not in the instructions

Example you can adapt

# ROLE
You are my strategic advisor. You know my operating priorities, recent decisions, and the shape of my week.

# JOB
Surface what I'm not seeing. When I bring you a plan, test it before you agree with it.

# FRAMES
Use my decision frames: capital allocation, talent, product velocity, brand. Default to those when evaluating any new question.

# OUTPUT
Memos: 1 page max, structured as Question, Options, Criteria, Recommendation, Open Questions, Reversibility.
Briefs: 1 page max with Top 3, replies drafted, heads up.

Scope · One task

Skill Instructions

Applies to: Whenever Claude loads that Skill. Claude reads only the Skill's description before deciding to load it, so you can also name the Skill in your request.

Where you set it: Wherever Skills are on. An Owner can share a Skill with the org as Available to install, Installed by default, or Required. Required controls installation. Claude still decides from the description whether to use it on a given task.

Put here

  • Repeatable workflows: 'turn this transcript into a follow-up email,' 'generate a board update from these KPIs,' 'triage these candidate notes into a scorecard'
  • A specific output template you want every time
  • Skill-specific tool use rules (e.g. 'always read the Drive doc first')
  • A description under 200 characters that says exactly when to use it. Anthropic's pages disagree on the limit (200 or 1,024), and 200 fits both.

Don't put here

  • Your role or operating frame - those are in Instructions for Claude
  • Project-specific context - that's in the Project the Skill is invoked from
  • A control that has to run on every task in a workflow. A Skill runs only when Claude picks it, so that rule goes in the Project's instructions.
  • Anything that changes per use - that goes in the prompt arguments, not the Skill instructions

Example you can adapt

# SKILL: meeting-prep-brief
INPUT: meeting topic + attendees (you'll get these via slash command).
ACTION: Pull my Strategic Advisor knowledge files. Generate a 1-page prep brief: meeting context, the 3 things to push on, the 1 question I should be asking.
OUTPUT: Plain markdown, no preamble. Save to my Drive folder called "Meeting prep" with the date in the filename.

Quick decision

Which scope

Would I want this in EVERY conversation I have with Claude?

→ Instructions for Claude.

Would I want this in every conversation EVERYONE on my team has?

→ Organization instructions - Team or Enterprise.

Would I want this only when I'm working on this specific role or domain?

→ Project instructions.

Is this a repeatable workflow I run on different inputs?

→ A Skill.

Examples

Project instructions examples for finance teams

Each rule names the work it applies to and gives the reason, so it fires where you meant it to and stays quiet on research. Fill in the brackets, then run the test set below before your team uses them.

Analyst learning Project

For first- and second-year analysts and associates building judgment on models and comps.

This Project helps analysts and associates build judgment on [comps, LBO and three-statement models, diligence findings, valuation views]. They need to defend every assumption under questioning.

When someone asks you to build one of these from scratch, or asks a judgment question about one, first ask for their approach in two to four bullets: main drivers, rough ranges, where they'd source each. Then give your version with the reasoning shown, say what their approach got right and where it breaks, and end with two questions a VP would ask.

Skip the question if they already gave an approach or say "skip". Formatting, cleanup, formulas, summarizing a document and finding a figure in a file always get a direct answer.

Tone: a helpful senior associate. Direct and brief.

Deal team workspace

One Project per live deal, shared with the deal team.

This Project supports the [Project Name] deal team evaluating [Target], a [business description]. The knowledge files hold the CIM, management presentation, QoE draft, and our model outputs as of [date].

Treat the knowledge files as the primary source and cite file and section. Label outside information and your own estimates as such. When files disagree (for example, the CIM vs. the QoE), show both figures, name the sources, and say which one our model uses.

For diligence questions, list findings in order of impact on value, then open questions for management, then items for the legal and accounting advisors.

Keep the working thesis in view: [one-line thesis]. When evidence cuts against it, say so plainly.

IC memo review

A review pass before a draft memo goes to the deal lead.

This Project reviews draft investment committee memos before they go to the deal lead. Act as a skeptical IC member who wants the deal to be good.

For each draft:
1. List every number and claim a reader can't trace to a source or stated assumption, with its location in the memo.
2. Name the three questions IC is most likely to ask that the memo doesn't answer.
3. Check that the returns math, sources and uses, and valuation figures agree across sections.
4. Flag any risk that is mentioned with no mitigant.

Give the author the list. Write suggested language only when the author asks, and only for that section.

LP update drafting

Quarterly letters, drafted for Investor Relations and Compliance to tie out.

This Project drafts quarterly LP updates for [Fund]. The knowledge files hold the prior four letters, our LP communications guidelines, and this quarter's portfolio data.

Match the voice and structure of prior letters. Use only figures from this quarter's data files, and list each figure with its source file at the end of the draft so Investor Relations and Compliance can tie it out. Describe performance factually; when a portfolio company underperformed, say so with the reason.

LP letters are checked for accuracy before release, so leave a placeholder in [brackets] for anything not in the data. Do not estimate here.

Before everyone gets it

Test before rollout

The common failure is a rule written for one case firing on everything else. Most of these prompts should pass straight through. The last three check that your rules fire when they should.

  1. Put the draft in a test Project first. Project instructions apply right away and only to the people in that Project.
  2. Run every prompt below in a fresh chat, before and after, and write down pass or fail.
  3. Pilot with three to five people across roles for a week.
  4. Before you change organization instructions, copy the current text somewhere safe. That copy is your rollback.
  5. Allow up to an hour after saving for the change to reach everyone, and run the set again whenever your default model changes.

Research stays open

“Size the US market for commercial HVAC services and name the five largest players.”

Pass: A full answer with labeled estimates and sources. No request for files.

Research stays open

“What EBITDA multiples are typical for vertical software buyouts right now?”

Pass: A range with sources or labeled estimates.

Diligence stays open

Paste a CIM excerpt: “What are the top five diligence risks?”

Pass: A ranked list citing the excerpt. No warning about the material.

Gaps don't stall the work

“Build a quick revenue bridge. I only have FY24 and FY25 totals.”

Pass: It proceeds, states its assumptions, and flags the missing data in one line.

Rote work stays open

“Reformat this comps table and sort by EV/EBITDA.”

Pass: It does it. No questions.

Learning rule fires (in the learning Project)

“Is a 12x entry multiple reasonable for this business?”

Pass: It asks for your view first.

Escape word works

“skip - is a 12x entry multiple reasonable for this business?”

Pass: A full answer, no question.

Sensitive-material rule fires when it should

“A friend at [public company] says they'll miss the quarter. How should we position our public holdings?”

Pass: One sentence of concern, a Compliance referral, and general analysis only.

Common mistakes

What to fix first

Every one of these is fixable in 10 minutes once you spot it.

Writing “stop and ask” rules into organization instructions

They run on every message from every person. A rule written for IC memos fires on a research question, and Claude starts hedging everything. Scope the rule to a Project and name the tasks it covers.

Putting the company brand voice in Instructions for Claude

Then it only applies to you. The marketing director writing the same email gets a different voice. Move it to organization instructions if you have Team or Enterprise.

Putting current-quarter priorities in Instructions for Claude

They change every quarter. You'll forget to update them. Put them in a Drive doc that your Strategic Advisor Project reads as a knowledge file - update the doc, the Project updates with it.

Stuffing long reference content into Project instructions

Project instructions are for behavior rules. Reference content goes in knowledge files (uploaded documents or linked Drive files). The Project reads the files when needed - much more capacity than instructions allow.

No personal instructions at all

Claude reverts to its default voice and behavior on every new chat - which means you keep retyping the same context. Even a 5-line file pays for itself the first week.

FAQ

Claude instructions questions

Which wins when organization and personal instructions conflict?

Organization instructions. Anthropic says Claude favors the organization-level instruction on a direct conflict, and that in rare edge cases behavior may vary. Personal instructions still apply to anything the organization instructions don't address.

Do Project instructions override organization instructions?

Anthropic doesn't publish an order for Project or Skill instructions. Write each layer so it doesn't contradict the ones above it, and test in a Project before relying on any order.

How long can Claude organization instructions be?

Up to 3,000 characters. They go out with every message from everyone in the organization, so shorter is better, and changes can take up to an hour to reach everyone.

Can I use Claude Projects on the free plan?

Yes. Free accounts can create up to five Projects. Sharing a Project with colleagues needs a Team or Enterprise plan.

Does a Required Skill always run?

No. Required means the Skill is installed for everyone. Claude still reads the description to decide whether to load it on a given task, so a rule that must run every time belongs in Project instructions.

Bonus

Get the Claude release brief

Claude ships changes faster than most instruction files get revised. The release brief is Nicole's read on each one, not Anthropic's: what to adopt, what to skip, and what it breaks in a setup you already wrote.

No spam. Unsubscribe anytime.

Occasional, never on a schedule, unsubscribe anytime. The generator above is free and does not ask for an email.

Setting this up across a firm?

The Custom Program walks your team through every scope live - organization, personal, Projects and Skills - built around your actual tools and roles. Private equity and capital management firms start with how we work with investment teams.