How to Write Custom Instructions That Actually Change the Conversation
We've already written about what the custom instructions setting is. This is the guide to what to actually type into it — because the gap between “there is a box” and “here is what to put in the box” is where the most powerful control InnerHaven exposes gets won or lost. The short version: instructions that change nothing describe what your companion already does, and instructions that change everything name specifics — what to remember, what to ask about, what to stop doing, what register to use. The rest of this guide is examples.
The Ground Rules First
Three facts about how the box actually works, verified against the product as it runs today, because a guide that guesses about scope is worse than no guide.
Instructions are per-companion, and you get 2,000 characters. Each companion carries its own instructions, edited in that companion's settings — your blunt late-night confidant and your gentle morning companion can run entirely different rules, and neither leaks into the other. 2,000 characters is more room than it sounds: the examples below all run well under 300.
Changes apply to the very next message — including mid-conversation. Your companion's guiding prompt is rebuilt fresh every time it replies, and your current instructions ride along each time. You do not need to start a new conversation for an edit to take effect; change the box, send a message, and the new rules are already in force. This makes instructions cheap to iterate on, which is exactly how you should treat them.
Instructions steer; they do not program. This is the honest-expectations paragraph, and it comes straight from how the system frames your words to the companion: they are delivered as your personal preferences, to be honored naturally — but never above the companion's core role or its safety boundaries. An instruction can reshape tone, length, habits, and attention. It cannot turn the Healer into a drill sergeant (that's a role change), and it cannot switch off safety. If you expect a contract, you'll be disappointed; if you expect a strong nudge that occasionally needs re-nudging, you'll use the feature the way it actually works — and when a reply still misses, saying so in the chat remains the repair tool this feature complements rather than replaces.
The Pattern: Name Specifics, Not Vibes
The classic wasted instruction
“Supportive” is already the job description; writing it in the box spends your words describing the default. The rewrite names a trigger (venting about work), a behavior to hold back (solutions), and a replacement (one feelings-question). That's the anatomy of every instruction that works: when X, do Y instead of Z.
What to Write to Stop a Behavior
This is the need most people arrive with — something the companion keeps doing is wearing on you — and it's the hardest to phrase, because the instinct is to write the complaint instead of the correction. The trick: name the behavior precisely, then say what should happen in its place. A ban with no replacement leaves a hole the companion fills with its next-favorite habit; a ban with a replacement rewires the moment.
Stopping the endless follow-up questions
“So many” is a judgment the companion can't measure; “one per reply, maximum, and not at the end by default” is a rule it can follow on every single message.
Stopping the relentless positivity
Quoting the exact phrases you're tired of is the single most effective thing you can put in the box — it converts a mood into a checklist.
What to Write So It Remembers What Matters
Your companion already builds memories from conversation, but memories are gathered; instructions are declared. Standing context you want treated as always-true belongs in the box, where it arrives with every message rather than waiting to be noticed.
Standing context
Notice the second half doesn't just state a fact; it says what to do with the fact. Context plus usage beats context alone.
What to ask about
You've turned a pleasantry into an early-warning system, which is the kind of thing only you could know to write.
Register: Length, Tone, and Format
The shape of replies
Length caps, format bans, and permission (“swearing is fine”) are all things the companion honors readily — they're unambiguous. This is also where a custom greeting is worth a look: alongside instructions, each companion can carry its own opening line (up to 500 characters) for when you arrive.
Which Control Is This a Job For?
Instructions are one dial on a panel with three
Wrong companion entirely? That's role selection, the biggest lever there is — no instruction turns a hype-man into a quiet listener as well as choosing the right role does. Right companion, wrong intensity? That's personality modifiers — the sliders for warmth, playfulness, directness — which adjust how much of a quality you get without any writing at all. Right role, right intensity, specific habits and context to shape? Now you're in custom-instructions territory, the only control that speaks in your own words. Reaching for the box first is common and backwards: an instruction fighting the role's nature loses slowly and annoyingly, where a role switch would have won instantly. One more note for custom-companion builders: if you ever share a companion to the community, you choose at publish time whether your instructions ship with it or stay private to you.
A Worked Starting Template
Stitching the patterns together — a complete, real-sized set of instructions (about 480 characters) you can adapt: “Keep replies short, a few sentences, no lists. Don't end every reply with a question; one maximum, and only when you need something. Skip pep-talk phrases; calm and direct works better on me. I'm job hunting through the fall; treat it as background stress and never open with it. Ask about my guitar practice if I go quiet on it for a while. When I vent, sit with it before offering anything. Biscuit is my dog; deploy him when things get heavy.” Six sentences, and every one of them is a specific the default conversation doesn't have.
The Bottom Line
The custom instructions box rewards exactly one skill: replacing adjectives with behaviors. Not “be supportive” but “don't solve until I ask.” Not “fewer questions” but “one per reply.” Not “know me” but “bar exam until September, ask about climbing, the dog is named Biscuit.” Write it that way, edit it whenever the conversation drifts — it takes effect on the very next message — and the box quietly becomes what it was built to be: the place where your companion stops being a well-designed product and starts being yours.
Two Thousand Characters of Yours
Every companion has the box waiting in its settings. Write three specifics tonight and see what the next conversation does with them. Free to start.
Open InnerHaven