Task statement 5.3 · Configuration and knowledge management

5.3 · System-Level Instructions

In Wrenfield Housing, the course example, the communications team's Tenant comms Project began with one instruction, "Follow the style guide in every letter." (3.1). Rules added one by one grew it into 40 lines of capitalised MUSTs and NEVERs. Claude now follows some too rigidly and switches between others that clash.

Why it's tested

The tempting fix when Claude gets a standing rule wrong is to add another rule, or to repeat the old one in capitals. Claude applies every line in every chat the instructions cover, so a louder rule spreads to cases it was never meant for, and a new rule that clashes with an old one leaves Claude to choose. The opposite habit fails too, because one vague line gives no purpose or reasons to work from.

KnowledgeWhat "system-level" means in the Claude apps

A system-level instruction is a standing instruction: you write it once, and Claude applies it in every chat it covers, not in one message. The Claude apps have three such layers. Project instructions apply to every chat in one Project, and members with "Can edit" access can change them (sharing is 5.1). Instructions for Claude, in your own Settings, apply to all your own chats, and only you see them. Organisation instructions apply to every conversation in a Team or Enterprise organisation. Only Owners and Primary Owners can set or view them, and they hold up to 3,000 characters. This rules out any answer that puts one team's rules in a layer that reaches everyone, or in one person's settings.

Interactive · which standing instructions reach a chat

Four layers of instructions on the left, each with who sets it, and an arrow to the chats it applies to on the right. A message applies to one chat. Project instructions apply to every chat in the Project, for every member. Instructions for Claude apply to all of one person's chats. Organisation instructions apply to every chat in the organisation and win a clash with personal instructions. A layer not set yet is dashed and labelled not set, and its scope is dashed and labelled not applied. LAYER · WHO SETS IT WHERE CLAUDE APPLIES IT Your message you, each time you ask the task at hand APPLIES TO THIS CHAT ONLY gone in the next new chat Project instructions members with Can edit e.g. the Tenant comms rules NOT SET APPLIES TO EVERY CHAT IN THE PROJECT for every member who uses it NOT APPLIED Instructions for Claude you, in your own Settings your working habits NOT SET APPLIES TO ALL YOUR OWN CHATS only you can see them NOT APPLIED Organisation instructions Owners · Team and Enterprise up to 3,000 characters NOT SET APPLIES TO EVERY CHAT IN THE ORGANISATION wins a clash with personal instructions NOT APPLIED The wider the layer, the more chats each line in it changes.

Add each layer and see which chats it reaches.

One message: a rule reaches Claude only if someone types it into this chat.

Project instructions: every chat in Tenant comms starts from the team's rules, for every colleague.

Instructions for Claude: your own habits reach all your chats, and nobody else sees them.

Organisation instructions: an Owner's standards reach every chat and win a clash with personal instructions.

Dashed and labelled NOT SET: the layer is empty. Solid: the chats that layer reaches.

When layers disagree, organisation instructions take precedence over individual ones, which still apply wherever the organisation's are silent. The help pages do not say how project instructions rank, and they warn that contradictory instructions may not be followed reliably. So give each layer one job: standards for everyone, your own habits, or one team's way of doing one kind of work.

More precise than the guide. The Claude apps never say "system-level instructions": their settings are organisation instructions, Instructions for Claude and project instructions. Anthropic's prompting guide uses system prompt, the standing instructions an app built on the API sends with every request (checked 2026-10-15). Exam answer: system-level instructions are the standing instructions that shape every chat in their scope, most often a Project's.

KnowledgeWhat makes standing instructions effective

Standing instructions are a prompt Claude reads in every chat they cover, so 1.1's parts still apply, but each line must now hold for requests nobody has thought of yet. Anthropic's prompting guide names three habits that matter most. A role, even in one sentence, focuses Claude's behaviour and tone. The reason behind each rule, as 1.1 showed, lets Claude apply it to cases the rule never named and skip cases it was never meant for. Telling Claude what to do works better than telling it what not to do. The prompting guide also warns that recent models follow instructions closely, so wording such as "CRITICAL: You MUST" can make Claude overdo a rule. This rules out any answer that fixes a misapplied rule by making it louder.

Build standing instructions in this order:

  1. Purpose, audience and role: whom Claude helps, with what work, for which readers.
  2. Rules with reasons: each rule on one line, with the "because" that shows where it applies.
  3. What to do: the behaviour you want, with a "never" only next to what to do instead.
  4. Format: the usual shape of the result, such as the opening and length.
  5. When to ask: ask, or say it does not know, when a fact is missing.
  6. Short and tested: only lines that change results, tried on real past requests in a new chat.

Step 5 matters because Anthropic's guidance on hallucinations says explicit permission to admit uncertainty can reduce false information. Step 6 is the help article's own advice for organisation instructions: concise, specific, and tested in a new conversation. One kind of line does not belong in instructions. The facts Claude works from go in the Project's knowledge files (managing them is 5.2):

How Claude works

Instructions

Role and audience, rules with reasons, format, when to ask.

Changes when: the team changes how it wants the work done.

What Claude works from

Knowledge files

Phone numbers, policies, prices and dates, kept in the document that owns them.

Changes when: the fact changes, in one place, with no rule rewritten.

SkillWorked example: the 40-line block Claude could not follow

The Tenant comms Project started with one line and later gained the annual report's three style decisions (3.4). Each wrong letter added a rule in capitals, until the block ran to 40 lines, the repairs line number included. Claude then applied some rules too rigidly and switched between others:

# before: 4 of the 40 lines
ALWAYS use bullet points.
Start every letter "Dear resident".
ALWAYS address the tenant by name!!!
Repairs line: [the number, typed in]
→ a condolence letter in bullet points; "Dear resident" in some letters, "Dear [NAME]" in others
# after: the whole block
You help Wrenfield's communications team draft letters and notices for tenants.
Readers are tenants of every age and reading level, so write in plain English.
Open with "Dear resident", because letters go to whole households.
Say "tenants", never "customers", because they rent their homes from us.
Keep sentences to 20 words or fewer, so letters are easy to read.
Give each figure its period and source, so readers can check it.
Use bullet points only for steps a tenant must take.
Take facts, including the repairs line number, from the style guide and handbook.
If a request needs a fact those files do not hold, ask me rather than guess.
→ three real past requests: the old block passed 1, the rewrite passed all 3

Each fault traces to a line in the block. "ALWAYS use bullet points" had no reason, so Claude used bullets even in a condolence letter, where they read as cold. Two rules told Claude to open letters in two ways, and the second invited staff to paste tenants' names, which Wrenfield's AI policy forbids. The rewrite gives a role and audience, keeps 3.4's three style decisions with a reason each, and moves the repairs line number back to the style guide. The team tested it in a new chat on three past requests, the Millbrook boiler letter, the spring newsletter and a communal cleaning notice, none holding tenant data.

Ask of each line: what kind of line is it, and where does it belong?
A rule for this team's workKeep it in the Project's instructions, phrased as what to do, with its reason.
A fact Claude works fromMove it to the knowledge files, where one edit updates it.
A rule for everyone, or only for youOrganisation instructions (an Owner sets them) or your own Instructions for Claude.
Takeaway
  • Standing instructions cover one Project, all your own chats, or the whole organisation, whose instructions win a clash with personal ones.
  • A short role, rules with reasons and plain "do this" wording beat a long list of capitalised MUSTs.
  • Test new instructions on real past requests in a new chat before colleagues rely on them.

check yourself

0 / 3 answered

A finance team of five, on their organisation's Team plan, wants every month-end report Claude drafts for them to use the same table layout and rounding rules. The marketing and sales teams also use Claude, with their own report formats. The rules change next quarter, and the team lead wants to update them in one place.What is the most effective place to put the rules?

Select one.

An HR team's Project instructions read: "NEVER give legal advice. NEVER discuss pay. ALWAYS be helpful." Line managers report that Claude refuses to explain the published pay bands and will not summarise the holiday policy, saying that might be legal advice. The pay bands are already in the Project's knowledge.Which TWO changes to the instructions would most directly fix this?

Select two.

A consultant at a firm on a Team plan added "Use American spelling." to their Instructions for Claude last month. Their chats, inside and outside Projects, still come back in British spelling, even with no files attached. Memory is turned off for them, and none of their Projects' instructions mention spelling.What is the most likely root cause?

Select one.

sources

  1. Claude Help Center, "Set organization instructions" (updated 2026-08-31): Owners and Primary Owners on Team and Enterprise plans; Organization settings; Claude follows them in every conversation across the organisation; 3,000 characters; they take precedence over individual instructions, which still apply where they are silent; only Owners and above can view them; contradictory instructions may not be followed reliably; keep them concise and specific, test them in a new conversation, review them regularly. Checked 2026-10-15.
  2. Claude Help Center, "Understanding Claude's personalization features": Instructions for Claude in Settings apply to all your conversations; project instructions apply only to chats in that Project. Checked 2026-10-15.
  3. Claude Help Center, "How can I create and manage projects?": Claude uses project instructions for all chats in the Project; members with "Can edit" can modify instructions and knowledge. Checked 2026-10-15.
  4. Claude Platform Docs, "Prompting best practices": a role in the system prompt focuses behaviour and tone, even in one sentence; context and motivation behind instructions; tell Claude what to do instead of what not to do; dial back aggressive language such as "CRITICAL: You MUST". Checked 2026-10-15.
  5. Claude Platform Docs, "Reduce hallucinations": explicitly allow Claude to say "I don't know" to reduce false information. Checked 2026-10-15.
  6. Wrenfield's 40-line instruction block, the condolence letter and the three test requests are this course's example.