What you will be able to do
- Decide whether a piece of content belongs in persistent instructions, Project knowledge, or the individual request
- Write standing rules for role, tone, constraints and format that are precise and short
- Spot conflicting or non-standing lines that make outputs uneven, and remove them
- Work through the steps of designing a Project's instructions in order, from listing the recurring requests to revising after a test
Key concept
Standing rule vs per-request content — A standing rule is a behaviour that should hold the same way in every conversation, such as tone, format, role or an off-limits topic. It belongs in persistent instructions. Anything that changes from one conversation to the next, such as a customer's case, this week's deadline or one person's client, belongs in the request that needs it.
1.Where persistent instructions live and how far they reach
Persistent instructions are text Claude receives before you type anything. They apply to every conversation in their scope, so you don't have to repeat them. When a team runs the same kind of request dozens of times a week, such as support replies, finance summaries or code reviews, these instructions are what keep the hundredth output looking like the first. They only work if they hold the right content, and the right content depends on how far each layer reaches.
Project instructions are the layer most teams start with. You set them on a Project and they cover every chat inside it. One example is telling Claude to use a more formal tone, or to answer from the point of view of a particular role or industry. How to create and share a Project is covered in Project configuration. This page is about what to write in the instructions.
| Layer | Set by | Applies to | When it conflicts |
|---|---|---|---|
| Organization instructions | Owners and Primary Owners (Team and Enterprise) | Every conversation of every person in the organization | Takes precedence over individual instructions |
| Individual user instructions | Each user | That user's conversations only | Still applies wherever organization instructions are silent |
| Project instructions | Members who can edit the Project | All chats within that Project | Scoped to the Project |
| CLAUDE.md (Claude Code) | You, committed to the repository for the team | Every session in that directory | If two instructions contradict, Claude may pick one arbitrarily |
The layers stack. Organization instructions are capped at 3,000 characters and go out with every message from everyone in the organization. Individual instructions fill in wherever the organization layer is silent. In Claude Code, CLAUDE.md does the same job for a codebase: Claude reads it automatically at the start of every session, so it never needs to be pasted into a prompt. The design rule is the same at every layer. The wider an instruction reaches, the more sure you need to be that it should apply every single time.
Checkpoint 1 of 8· Match them up
Match each layer of persistent instructions to how far it reaches.
Tap a term, then the definition that fits it.
Each layer reaches a different distance: the whole organization, one person, one Project's chats, or one codebase's sessions. The wider an instruction reaches, the more sure you need to be that it should apply every time.
“Organization instructions let owners of Team and Enterprise plans set custom instructions that Claude follows in every conversation across your organization.”Source: support.claude.com
2.Dividing the work: standing rules vs the request
The test for whether a line belongs in persistent instructions is simple: should this be true of every output, whatever is being asked? Tone, greeting, sign-off, output format, the role Claude plays and topics it must not touch all pass. Whatever this chat is about fails: a customer's message, case details, a deadline, a reference to last quarter. That material goes in the request, where it applies to exactly one conversation.
Each wrong placement fails for its own reason. Project instructions are used for all chats in the Project, so swapping one customer's details in and out turns a shared setting into a per-chat chore. Someone will forget to swap, and the next reply will cite the wrong case. Project knowledge has the same reach: anything uploaded there is used across every chat, so it suits reference material all chats need, such as a policy handbook, not one customer's file. Putting the standing rules in each request and the case details in the instructions inverts the design. The parts that should never change now depend on someone remembering to paste them, and the part that changes every time sits where it applies everywhere.
| Content | Same in every chat? | Place it in |
|---|---|---|
| Greeting, warm-but-concise tone, sign-off | Yes | Project instructions |
| Length and list formatting rules | Yes | Project instructions |
| Refund policy every reply may need to cite | Yes, as reference | Project knowledge |
| This customer's message and case history | No | The request for that chat |
| A one-off note such as a template change starting Monday | No, it expires | The requests it affects |
Non-standing lines creep into instructions because they are useful when written. Take a finance Project whose instructions have grown to 45 lines. Mixed in with the formatting rules are a reminder about last quarter's audit, a note to use a new template starting Monday, and the name of one analyst's current client. Each of these helped once. Each is now applied to every chat, including chats it has nothing to do with, and outputs drift. The fix is to sort by the test. The formatting and tone rules stay. The audit reminder, the template note and the client name move into the requests that need them.
Checkpoint 2 of 8· Check yourself
A finance team's Project instructions have grown to 45 lines. Alongside the formatting and tone rules are a reminder about last quarter's audit, a note to use a new template starting Monday, and one analyst's current client name. Outputs have started to drift. Which fix restores uniform outputs?
Project instructions apply to every chat in the Project, so case-specific and time-bound lines affect chats they have nothing to do with. Reordering or repeating rules doesn't change that, and Project knowledge reaches every chat too. Only the request is scoped to a single conversation.
“Claude will use these instructions for all the chats within the project.”Source: support.claude.com
Checkpoint 3 of 8· Exam question
A finance team's Project instructions have grown to 45 lines. They include formatting rules, a reminder about last quarter's audit, a note to 'use the new template starting Monday', and the name of one analyst's current client. Outputs have become uneven. What is the most appropriate fix?
Correct answer: C — Keep the formatting and tone rules in the Project instructions, and supply the audit reminder, template note and client name in the requests that need them.
- A. Incorrect: headings and a priority order reorganise the 45 lines but leave the time-bound template note and one analyst's client name in instructions that shape every output the team produces, so the unevenness remains.
- B. Incorrect: restating the formatting rules lengthens the instructions further and does nothing about the dated note, the audit reminder or the single client name, which still do not belong in standing instructions.
- C. Correct: standing instructions should hold durable rules that apply to all of the team's work, such as format and tone. Dated, topical or person-specific items are better given in the requests that need them.
- D. Incorrect: this removes the audit and template context that some requests still need, while leaving one analyst's client name in instructions that apply to the whole team's outputs, so the person-specific item is still in the wrong place.
Checkpoint 4 of 8· Exam question
A customer-support team uses a Claude Project to draft replies. Every drafted reply must use the same greeting, the same warm-but-concise tone and the same sign-off, while each chat concerns a different customer. Which split of content is the best design?
Correct answer: B — Put the greeting, tone and sign-off rules in the Project instructions, and paste each customer's message and case details into the request for that chat.
- A. Incorrect: case details belong to one customer, but each chat concerns a different one. The instructions would need editing before each chat, and a missed edit would carry one customer's details into another customer's draft.
- B. Correct: rules that must hold in every reply are standing, so they belong in Project instructions, while the customer message changes each time and belongs in the request. This keeps outputs uniform without retyping rules.
- C. Incorrect: Project knowledge is shared by all chats in the Project, so different customers' case details would be available side by side in each chat, even though each chat concerns a single customer. Case details belong in that chat's own request.
- D. Incorrect: this reverses the split. Standing rules retyped into each request can drift between replies, and one customer's case stored in the Project instructions would apply to chats about other customers.
3.Writing standing rules: precise, not bloated
A good standing rule is concrete enough that two people would read it the same way. "Be professional" leaves Claude to guess. "Respond in formal English. Use active voice. Avoid contractions and emojis." leaves nothing to guess. Format rules work the same way: "Prefer concise responses under 300 words. Use bullet points for lists with three or more items." gives a limit Claude can meet every time. A useful mental model is a brilliant new employee who lacks context on your norms. Spell out what that person couldn't know. The general techniques are in prompt construction. Here they are applied to text that runs on every request.
Standing rules usually cover four things:
- Role. One sentence focuses behaviour and tone, for example a healthcare-claims context where "claims" means insurance claims, not legal claims.
- Tone. Name the register and what it excludes.
- Constraints. Off-limits topics and referrals, such as sending HR policy questions to the HR inbox instead of giving policy advice, plus data-handling rules like leaving customer names and account numbers out of responses.
- Format. Length, structure and the template every output follows.
Giving the reason behind a rule makes it stronger. "Never use ellipses" is weaker than explaining that the text will be read aloud by a text-to-speech engine that can't pronounce them, because Claude can generalize from the explanation to cases the rule didn't name.
Length works against all of this. Every line is loaded into every conversation, and in Claude Code's guidance for CLAUDE.md, longer files consume more context and reduce adherence. That guidance targets under roughly 200 lines and leaves out aspirational rules the team does not actually follow. Headers and bullets help Claude follow the rules that are there. They don't fix bloat. Sorting 45 mixed lines into ranked sections, or repeating the format rules at the top and bottom, leaves the stale and person-specific lines in place, still applying to every chat. The cure for bloat is to remove the lines, not to rearrange them.
Checkpoint 5 of 8· Check yourself
Replies in a shared Project vary in style from chat to chat. Its only tone instruction reads "Be professional." Which replacement is most likely to make every reply read the same way?
A standing rule should be concrete enough that two people would read it the same way. Naming the register and what it excludes leaves nothing to guess. Appeals to reputation, mirroring each customer, or repeating a vague rule for emphasis still leave Claude guessing, so replies keep varying.
“Instead of vague instructions like “be professional,” give concrete direction such as “Respond in formal English. Don’t use contractions, slang, or emojis.””Source: support.claude.com
4.Conflicting instructions and which layer wins
Two standing rules that contradict each other don't settle into a sensible compromise. Claude may follow neither one reliably, or pick one arbitrarily, and the choice can differ from chat to chat. That is the opposite of what persistent instructions are for. Conflicts tend to build up as different people add lines over months, so read the instructions as a whole, not one line at a time. Where the two rules really compete, write down which one wins, for example "keep replies under 150 words unless the user asks for a walkthrough".
Between layers, the precedence is defined. When organization and individual instructions directly contradict, the organization instruction wins: if the organization says always respond in formal English and a user's own instructions ask for a casual tone, Claude responds formally. Individual instructions still apply to anything the organization layer doesn't address. Even this ordering relies on prompt-level instructions, and behaviour may vary in rare edge cases of direct contradiction, so test the combination instead of assuming it.
Checkpoint 6 of 8· Check yourself
Over several months, two editors added these lines to a shared Project's instructions: "Keep replies under 150 words" and "Include a full step-by-step walkthrough." Some replies are now short and others are long. Which change makes replies consistent?
When two standing rules contradict, Claude may follow neither reliably, and its choice can differ from chat to chat. Emphasis or moving one rule to Project knowledge leaves the conflict in place, because knowledge also reaches every chat. Stating which rule wins removes the conflict.
“Avoid conflicting instructions. If your organization instructions contradict each other, Claude may not follow either one reliably.”Source: support.claude.com
5.Designing instructions for a recurring task, step by step
Uniform outputs come from planning for the requests before writing anything. Instructions written from a single example fit that example. Instructions written from the full range of requests the Project will see hold up against all of them. The process runs from understanding the work to refining the text:
- List the recurring requests the Project will handle and who reads each output, such as a customer, an auditor or an internal team.
- Identify the constraints that must hold every time: tone, format, off-limits topics, data-handling rules.
- Draft concise instructions that hold only those standing rules and leave case-specific details to requests.
- Test the draft against both typical and unusual requests, starting a new conversation for each, and compare the outputs for uniformity.
- Revise the instructions to close any gaps or conflicts the comparison revealed.
Testing only the typical case is the shortcut that shows up later as uneven output. Try several different kinds of questions to confirm the instructions work across topics, and include edge cases, the same way good examples cover them so Claude doesn't pick up an unintended pattern. If the unusual request gets a different greeting or breaks the format, the instructions have a gap, and step 5 is where you close it. Keeping instructions current as the work changes over months is covered in configuration lifecycle.
Checkpoint 7 of 8· Check yourself
Before rolling out a support Project, a team drafted its instructions from one typical refund question. They checked the draft on that same question, and the reply looked right. Which step is still missing before they can trust the instructions to keep outputs uniform?
Instructions drafted and checked against a single example only fit that example. Testing typical and unusual requests, each in a new conversation, and comparing the outputs shows the gaps that step 5 then closes. Repeating the one case, moving the rules into requests, or sharing one example through Project knowledge leaves those gaps hidden.
“Test your instructions. After saving, start a new conversation to verify Claude is following your instructions. Try a few different types of questions”Source: support.claude.com
Checkpoint 8 of 8· Exam question
A new team lead is designing Project instructions for a recurring task and wants the design process itself to be sound. Which sequence works from understanding the work to refining the result?
Correct answer: A — List recurring requests and their readers → Identify constraints that must always hold → Draft concise standing instructions → Test draft on typical and unusual requests → Revise to fix revealed gaps or conflicts
- A. Instructions can only anticipate real use once the requests and readers are known, so that comes first. The constraints that hold across those uses come next, then a concise draft, then testing with output comparison, and finally revision of the gaps and conflicts the testing revealed.
- B. Drafting before the constraints are identified means the standing rules (tone, format, off-limits topics) get invented ad hoc while writing, so the draft can miss or contradict what must hold for every request, and the constraints only appear after the wording is already set.
- C. Identifying constraints before listing the recurring requests and readers means the rules are chosen without knowing which uses and audiences they must cover, so they may fit the wrong cases and leave real ones unaddressed.
- D. Testing a draft before the constraints are identified means the comparison has no agreed standing rules to check uniformity against. The constraints would then be added only after testing, which leaves the earlier test results unreliable.
Sources
(7)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Claude will use these instructions for all the chats within the project.”
↩︎ Where persistent instructions live and how far they reach“Anything you upload to this space will be used across all of your chats within that project.”
↩︎ Dividing the work: standing rules vs the request“Anything you upload to this space will be used across all of your chats within that project.”
↩︎ Exam trap 1“Claude will use these instructions for all the chats within the project.”
↩︎ Checkpoint - 2.
“For example, instructing Claude to use a more formal tone or answer questions from the perspective of a specific role or industry.”
↩︎ Where persistent instructions live and how far they reach - 3.
“Organization instructions are included in every message sent by everyone in your organization”
↩︎ Where persistent instructions live and how far they reach“Focus on consistent behaviors.”
↩︎ Dividing the work: standing rules vs the request“Respond in formal English. Use active voice. Avoid contractions and emojis.”
↩︎ Writing standing rules: precise, not bloated“Prefer concise responses under 300 words. Use bullet points for lists with three or more items.”
↩︎ Writing standing rules: precise, not bloated“If your organization instructions contradict each other, Claude may not follow either one reliably.”
↩︎ Conflicting instructions and which layer wins“When both are set, organization instructions take precedence.”
↩︎ Conflicting instructions and which layer wins“Try a few different types of questions to confirm the instructions work across a range of topics.”
↩︎ Designing instructions for a recurring task, step by step“Organization instructions work best for instructions that should apply uniformly across every conversation”
↩︎ Key concept“If your organization instructions contradict each other, Claude may not follow either one reliably.”
↩︎ Exam trap 3“Organization instructions let owners of Team and Enterprise plans set custom instructions that Claude follows in every conversation across your organization.”
↩︎ Checkpoint“Instead of vague instructions like “be professional,” give concrete direction such as “Respond in formal English. Don’t use contractions, slang, or emojis.””
↩︎ Checkpoint“Avoid conflicting instructions. If your organization instructions contradict each other, Claude may not follow either one reliably.”
↩︎ Checkpoint“Test your instructions. After saving, start a new conversation to verify Claude is following your instructions. Try a few different types of questions”
↩︎ Checkpoint - 4.https://support.claude.com/en/articles/14553240-give-claude-context-claude-md-and-better-promptsOfficial docs
“CLAUDE.md is a plain markdown file that Claude automatically reads at the start of every session in that directory.”
↩︎ Where persistent instructions live and how far they reach“Aspirational rules the team does not actually follow.”
↩︎ Writing standing rules: precise, not bloated - 5.https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-featuresOfficial docs
“These instructions only apply to chats within that project.”
↩︎ Dividing the work: standing rules vs the request - 6.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
↩︎ Writing standing rules: precise, not bloated“Claude can generalize from the explanation.”
↩︎ Writing standing rules: precise, not bloated“Diverse: Cover edge cases”
↩︎ Designing instructions for a recurring task, step by step - 7.https://code.claude.com/docs/en/memoryOfficial docs
“Longer files consume more context and reduce adherence.”
↩︎ Writing standing rules: precise, not bloated“if two instructions contradict each other, Claude may pick one arbitrarily.”
↩︎ Conflicting instructions and which layer wins“Longer files consume more context and reduce adherence.”
↩︎ Exam trap 2