Instructions

Instructions: rules you set for every coding agent

Instructions are the standing rules you write yourself, once, in Settings. Context Mode adds them to every request from Claude Code and Codex, on every machine. Memory is different: it is learned from your sessions over time.

Key facts

What it is: One text you write, 4,096 characters at most, that tells every agent how to answer you.
Who sets it: You, explicitly, in Settings. Nothing is added to it from your sessions.
Clients: Claude Code and Codex, through Context Mode. Sessions that go direct to Anthropic or OpenAI do not get it.
Cost: Read from the provider's cache after the first request, so it costs almost nothing per turn.
Limits: It changes how the agent writes, not what it does. A longer text is refused, not cut.

Instructions or Memory

	Instructions	Memory
Who writes it	You, in Settings	Context Mode, from your sessions
When it changes	Only when you save a new text	Over time, as facts change
What it holds	How the agent should answer you	Facts about you and your projects
Example	"Keep answers short. No wall of text."	"production is deployed to Render, Oregon"
History	None. The current text is the rule.	Each fact keeps its old values with dates

Use Instructions for a rule you want followed from the first request. Use Memory for what the agents should learn as you work.

What it does

Instructions change how the agent writes, not what it does. They add no task. Context Mode marks them as a standing rule, and says that your current message wins if the two disagree. So "Answer in one line" in a message beats Instructions that ask for detail.

Where to set it

Open Settings in the console.
Find the Instructions card. It applies to every agent on your account.
Type or paste your text and click Save.

length4,096 characters at most. A longer text is refused, not cut.

scopeyour account: Claude Code and Codex, on every machine

turn offclear the box and save. The gateway then adds nothing. On an account that has never saved Instructions, the box shows the default instructions below.

Default instructions

This is the default. An account that has never saved Instructions gets this text, and the Instructions card in Settings shows it. Edit it and save to make it yours:

Talk to me in ASD-STE100 Simplified Technical English.
Writing rules, from Orwell, 1946. These govern prose: docs, PR text, messages. Never touch code or technical terms; swap in everyday words only where precision survives.

Never use a metaphor, simile or other figure of speech which you are used to seeing in print.
Never use a long word where a short one will do.
If it is possible to cut a word out, always cut it out.
Never use the passive where you can use the active.
Never use a foreign phrase, a scientific word or a jargon word if you can think of an everyday English equivalent.
When responding to me and creating text for me: Keep language concise and short, a 15yo should always be able to understand what you're writing. Avoid wall of text.

It is 756 characters, well under the limit. The first line asks for ASD-STE100, a controlled English made for aerospace maintenance manuals. The next line limits the rules to prose, so code and technical terms stay exact. Five lines are George Orwell's rules from 1946. The last line asks for short text.

This full text is the built-in default. An account that saved its own Instructions keeps it. An account that cleared the box and saved gets none.

Instructions asking for short answers were on in every benchmark run through the gateway. Context Saving says how that affects the results.

How it is applied

Every request. The gateway adds the Instructions to each request in a conversation, not only the first. The client's files and settings do not change.
At the start of the conversation, not in the system prompt. The Instructions goes at the start of the first message. The client's system prompt is never edited.
From the provider's cache. The Instructions sits in the part of the request that the provider caches. After the first request, it is charged at the cache read price, not as new input on every turn.
Stable during a conversation. To keep that cache warm, the gateway holds the start of a conversation unchanged. New Instructions reaches a running conversation after 65 minutes without a request, and every new conversation at once.
Only where people read it. Claude Code's helper calls, such as a session title or a page summary, get no Instructions.
Added once. If a client sends the Instructions block back, the gateway does not add a second copy.

Clients

Claude Code and Codex, through the gateway. Sessions that go direct to Anthropic or OpenAI do not get the Instructions. Connect each machine with npx @context-mode/cli, as in the quick start.

Compared with other approaches

Approach	Where it lives	Clients	Who sets it
Claude Code output styles	Built-in styles, such as Concise, or a custom style file. "To make a style your default across projects, set `outputStyle` in `~/.claude/settings.json`."	Claude Code	You, per user; a project's own settings take precedence. Admins can ship styles in the managed settings folder.
Claude Code CLAUDE.md	Files in the repo, in your home folder, or deployed by an admin	Claude Code	You, your team through git, or IT through a managed policy file that users cannot exclude
Codex AGENTS.md	Files you write	Codex	You. OpenAI's memories page says required team guidance belongs in AGENTS.md.
Cursor rules	Files in the project, plus team rules	Cursor	You, or a team admin for required team rules
Context Mode Instructions	The gateway, one text per account	Claude Code and Codex	You, once, in Settings

Where they are stronger. Claude Code output styles are first-party, set per user across projects, and need no third party. Instruction files are free, first-party and need no third party too. They can differ per project, and they can hold far more than writing rules: build steps, conventions and project facts. Cursor team rules let an admin require one rule for a whole team; Context Mode Instructions belong to one account.

Where Context Mode differs. One text covers Claude Code and Codex, on every machine, with no file to copy into each repository. It is marked as a rule about wording only, so it does not add work.

Compared with other tools: see the landscape.

Quick start → Active Context Memory