What you will be able to do
- Explain how organization-level and individual instructions interact, and which one wins when they conflict
- Turn a vague behavioural wish into a concrete, motivated instruction Claude can pattern-match
- Use role and explicit scope so an instruction lands where you meant it to
- Keep an instruction set short, internally consistent, and inside the character limit
- Verify a newly saved instruction set before relying on it, and recognise letter-over-spirit failure
Key concept
System-level instruction — A durable instruction that sits above any single chat: once saved, Claude applies it to every conversation in its scope, rather than to the one message you happen to be typing.
1.Two levels, and which one wins
Before you can write a good system-level instruction you need to know who else is writing one. Claude supports two levels: instructions set for the whole organization, and instructions each person sets for their own conversations. They are not alternatives — both are live at the same time, and the interesting question is what happens where they disagree.
The organization level wins: "When both are set, organization instructions take precedence." That is not a total override, though. The personal instruction is not discarded — it simply loses on the point of conflict. "Individual user instructions still apply for anything not addressed by the organization instructions." So a personal preference for bullet-point summaries survives happily underneath an organization-wide formality rule.
| Level | Set by | Scope | Visibility |
|---|---|---|---|
| Organization instructions | Owners and Primary Owners | All people in your organization, every conversation | Only Owners and above can view or edit |
| Individual user instructions | Each user | That user's conversations only | Only that user can view or edit |
2.Specific beats well-meaning, and reasons beat rules
Most instruction sets fail not because they are in the wrong place but because they say nothing checkable. "Be professional" describes a feeling, not a behaviour. The guidance is blunt: "Instead of vague instructions like "be professional," provide concrete direction such as "Respond in formal English." Add what that actually forbids — no contractions, no slang, no emojis — and you have something Claude can pattern-match on every turn.
The same principle runs through Anthropic's general prompting guidance: "Claude responds well to clear, explicit instructions." The useful mental model is a brilliant new employee who has no idea what your team's norms are. If you want above-and-beyond behaviour, ask for it in words rather than hoping it is inferred.
| Less effective | More effective |
|---|---|
| Create an analytics dashboard | Create an analytics dashboard. Include as many relevant features and interactions as possible. |
| NEVER use ellipses | Your response will be read aloud by a text-to-speech engine, so never use ellipses. |
Because the reason generalises. "Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important" helps Claude understand the goal, so it also avoids the neighbouring things a speech engine would mangle — not just the literal character you banned.
There is a cheap test for whether an instruction is specific enough, and it does not involve Claude at all: "Show your prompt to a colleague with minimal context on the task and ask them to follow it." If they would be confused, so will Claude.
3.Set a role, and say how far it reaches
Specificity tells Claude what to produce. A role tells it from where. "Setting a role in the system prompt focuses Claude's behavior and tone for your use case" — and a single sentence is enough to shift the output. If you want drafts written from the perspective of the person who will assess them, name that perspective in the durable instruction rather than re-typing it into every chat.
Domain context belongs at this level too, because it is the shared vocabulary nobody should have to re-explain: "Our team works in healthcare claims processing." Once that is a standing instruction, an ambiguous word like "claims" resolves the way your team means it, in every conversation.
The counterpart is scope. Recent models read instructions literally and do not silently widen them: "If you need Claude to apply an instruction broadly, state the scope explicitly". The mirror-image risk is a durable instruction whose scope is wider than you intended — an audience rule written for one kind of output quietly steering every other kind of output produced in the same place.
A communications team wants every press-release draft produced in a shared project to come out in the company's house voice, without anyone having to explain the voice again in each chat. What should the team do first?
Correct answer: A — Write down the house voice rules and audience the team already applies, then put them in the project instructions.
- A. Correct. The durable direction has to exist as agreed wording before it can be configured; defining what good looks like comes first.
- B. This is the repeated one-off prompting the team is trying to escape, and it produces different results per person.
- C. Samples are useful background, but they do not state the rules, and inference from mixed past work is unreliable.
- D. This guarantees inconsistency, which is the opposite of the shared house voice the team asked for.
4.Short, consistent, and not a way around safety
System-level instructions are not free. "Organization instructions are included in every message sent by everyone in your organization", so length is a standing cost paid by every conversation, and the hard ceiling is real: "The maximum length is 3,000 characters." Direct, specific guidance beats a lengthy explanation of your reasoning.
The bigger danger in a long set is self-contradiction. "If your organization instructions contradict each other, Claude may not follow either one reliably." Note the failure mode: not "the first rule wins" and not "the stricter rule wins", but unreliability on both. Review the set as a whole, not rule by rule.
An operations lead runs a project whose instructions describe the team's standard weekly status report. This week only, a board member has asked for a one-page version. What is the best way to handle the request?
Correct answer: B — Ask for the one-page version in that chat, leaving the project instructions describing the normal report.
- A. Knowledge is background material, not direction, and this still makes a temporary need permanent.
- B. Correct. A one-off need belongs in the individual request; durable instructions should keep describing the standing format.
- C. Baking a one-off requirement into durable direction changes every future report for a single exception.
- D. A whole parallel workspace is heavy maintenance for a single request, and the two projects will drift apart.
5.Saving is not the last step
An instruction set is a hypothesis about behaviour until you have watched it behave. The prescribed check is small and concrete: "After saving, start a new conversation to verify Claude is following your instructions." A new conversation matters, because an open one was started under the old configuration.
One probe is not enough either. "Try a few different types of questions to confirm the instructions work across a range of topics." That is how you catch a rule that holds on the case you had in mind and collapses on the case you did not.
Build a little patience into the check: "Changes may take up to an hour to take effect across Claude products." A first test that shows the old behaviour is not automatically a badly written instruction.
A project's instructions say "always write in plain language for non-specialists" and, further down, "use the full technical terminology our engineers expect". Team members report that outputs vary widely. What is the greatest concern with these instructions as written?
Correct answer: C — The two rules conflict, so Claude has no reliable basis for choosing a register and results vary between chats.
- A. Instruction wording does not trigger any reprocessing of uploaded documents.
- B. A predictable midpoint is not what happens; the observed symptom is inconsistency, not a stable compromise.
- C. Correct. Contradictory house rules are the direct cause: neither rule can be satisfied without breaking the other.
- D. There is no such rule that only the opening line of project instructions is read.
Sources1
6.When it obeys you and still gets it wrong
The last skill is diagnostic. Instruction-following is pattern continuation, which is why the reliable zone is narrow and concrete: "Capability zone: short, concrete, verifiable instructions." Format specs, length limits and explicit roles land precisely. Abstract or ambiguous direction lands somewhere in the gap between what you wrote and what you meant.
The wrong fix is emphasis: "Repeating the instruction with more force won't close the gap." Shouting a rule that was already followed changes nothing, because the rule was never the problem. The right fix is to state the outcome you are after: "When an instruction is followed literally but uselessly, restate the goal." Written into a durable instruction, that means pairing the format rule with the purpose it serves — which is the same move as attaching a reason, arriving from the other direction.
Sources5
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.A user's own instruction overrides the organization instruction, because it is closer to the conversation.Why is that wrong?
Precedence runs the other way: the organization level wins on any point of direct conflict, and the personal instruction survives only for whatever the organization instructions leave unaddressed.
Covered in Two levels, and which one wins
2.A carefully worded organization-level instruction can relax Claude's safety guidelines for internal, trusted use.Why is that wrong?
No level of instruction can switch off built-in safety guidelines or content policies; instructions that conflict with Claude's core training are not followed at all, so the wording is beside the point.
3.If two standing instructions conflict, Claude will just pick the stricter or more recent one, so adding rules is always safe.Why is that wrong?
Contradiction degrades both rules rather than resolving in favour of one, so a longer set can behave worse than a short one — and every added character rides along on every message.
4.When output still misses the point despite following the instruction, the fix is to restate the same rule more emphatically.Why is that wrong?
That is a letter-over-spirit failure: the instruction was honoured and the intent was not. Force adds nothing; stating the underlying goal alongside the instruction is what closes the gap.
Covered in When it obeys you and still gets it wrong
5.A standing instruction naming one audience is harmless for other work done in the same place, because Claude will infer when it does not apply.Why is that wrong?
Claude reads instructions literally and does not infer requests you did not make, so an audience rule keeps applying until you state its scope explicitly.
Covered in Set a role, and say how far it reaches
Practise it for real
Write and validate one organization-level instruction end to end, from vague wish to verified behaviour.
1.Write down the behaviour you want as a vague sentence, then rewrite it as a concrete rule with its reason attached — for example a formatting standard such as preferring concise responses under 300 words.
Why: Concrete direction outperforms an adjective, and the attached reason lets Claude generalise to nearby cases you did not enumerate.
You should see: A rule a colleague with no context could follow without asking you a question.
2.Go to Organization settings > Organization and access and find the Organization instructions section.
Why: This is where instructions that apply to every conversation in the organization are stored; the field is visible only to Owners and above.
You should see: A text area for the instruction set.
3.Paste your rule, check the whole set stays within 3,000 characters and that nothing already in the box contradicts it, then click Save changes.
Why: The set is sent with every message, and mutually contradictory rules make Claude unreliable on both.
You should see: The instructions save; the set reads as one consistent voice rather than a pile of rules.
4.Start a new conversation and ask two or three differently shaped questions that should each trigger the rule.
Why: A single probe can pass by luck; varied topics show whether the rule generalises.
You should see: The rule holds across all of them — though allow up to an hour for the change to take effect across Claude products before treating a miss as a wording problem.
5.If a response obeys the rule but still misses what you wanted, add the goal beside the rule rather than strengthening its wording, and re-test.
Why: Letter-over-spirit failures are not fixed by emphasis.
You should see: Output that serves the purpose, not just the format.
Stuck? Get a nudge
If you cannot state the reason for a rule in one clause, it is probably two rules — or one you do not actually need.
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“When both are set, organization instructions take precedence.”
↩︎ Two levels, and which one wins“Instruction prioritization relies on prompt-level instructions.”
↩︎ Two levels, and which one wins“Our team works in healthcare claims processing.”
↩︎ Set a role, and say how far it reaches“Organization instructions are included in every message sent by everyone in your organization”
↩︎ Short, consistent, and not a way around safety“The maximum length is 3,000 characters.”
↩︎ Short, consistent, and not a way around safety“If your organization instructions contradict each other, Claude may not follow either one reliably.”
↩︎ Short, consistent, and not a way around safety“After saving, start a new conversation to verify Claude is following your instructions.”
↩︎ Saving is not the last step“Try a few different types of questions to confirm the instructions work across a range of topics.”
↩︎ Saving is not the last step“Changes may take up to an hour to take effect across Claude products.”
↩︎ Saving is not the last step“Organization instructions let owners of Team and Enterprise plans set custom instructions that Claude follows in every conversation across your organization.”
↩︎ Key concept“When both are set, organization instructions take precedence.”
↩︎ Exam trap 1“If your organization instructions contradict each other, Claude may not follow either one reliably.”
↩︎ Exam trap 3 - 2.https://support.claude.com/en/articles/14503675-organization-instructions-in-claude-for-governmentOfficial docs
“Individual user instructions still apply for anything not addressed by the organization instructions.”
↩︎ Two levels, and which one wins“Instead of vague instructions like "be professional," provide concrete direction such as "Respond in formal English.”
↩︎ Specific beats well-meaning, and reasons beat rules“Organization instructions cannot disable Claude's built-in safety guidelines or content policies.”
↩︎ Short, consistent, and not a way around safety“Organization instructions cannot disable Claude's built-in safety guidelines or content policies.”
↩︎ Exam trap 2 - 3.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Claude responds well to clear, explicit instructions.”
↩︎ Specific beats well-meaning, and reasons beat rules“Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important”
↩︎ Specific beats well-meaning, and reasons beat rules“Show your prompt to a colleague with minimal context on the task and ask them to follow it.”
↩︎ Specific beats well-meaning, and reasons beat rules“Setting a role in the system prompt focuses Claude's behavior and tone for your use case.”
↩︎ Set a role, and say how far it reaches - 4.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-sonnet-5Official docs
“If you need Claude to apply an instruction broadly, state the scope explicitly”
↩︎ Set a role, and say how far it reaches“If you need Claude to apply an instruction broadly, state the scope explicitly”
↩︎ Exam trap 5 - 5.
“Capability zone: short, concrete, verifiable instructions.”
↩︎ When it obeys you and still gets it wrong“Repeating the instruction with more force won't close the gap.”
↩︎ When it obeys you and still gets it wrong“When an instruction is followed literally but uselessly, restate the goal.”
↩︎ When it obeys you and still gets it wrong“When an instruction is followed literally but uselessly, restate the goal.”
↩︎ Exam trap 4