> ## Documentation Index
> Fetch the complete documentation index at: https://shapesinc-4644c49f-cursor-docs-product-refresh-1067.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Designing AI With Social Intelligence
> Chapter 4 of the Make Shapes guide: the craft of behaving well in a group chat. When to speak, when to stay quiet, group vs DM, and tuning Free Will to read the chat.
*Chapter 4 of the [Make Shapes](/make-shapes) guide.*
Putting an AI into a group chat is easy. Making it *good* in one is the real skill, and it's what Shapes is built to let you get right. This is the chapter that sets your Shape apart.
A solo AI only has to answer the person typing. A Shape in a group chat has a harder job. It has to know **when to speak and when to stay quiet**, who it's talking to, what already happened, and how to share the floor with other Shapes and people. That's *social intelligence*: the skill of reading a chat. You design it; it isn't a switch you flip.
**What you'll learn:** how a Shape decides when to talk, how to tune that with Free Will, and how to make a Shape funny, calm, or helpful without being the bot that won't shut up.
## The core problem: presence without spam
Every annoying group-chat bot makes the same mistake. It treats every message as a cue to respond, and the result is a wall of AI text that drowns out the people.
The opposite failure is just as bad: a Shape so passive it never adds anything.
Good social intelligence lives in between. You want a Shape that's **present**. It picks up the thread, lands a reply at the right moment, then gets out of the way. Almost every control below helps you hit that balance.
## How a Shape decides to speak
When someone posts in a chat, here's how Shapes figures out who replies:
```mermaid theme={null}
flowchart TD
A[Someone posts a message] --> B{Is a Shape mentioned?}
B -- Yes --> C[Those Shapes reply, in turn]
B -- No --> D{Any Shapes activated?}
D -- Yes --> E[Activated Shapes reply in turn, up to five]
D -- No --> F{Free Will triggers?}
F -- "mention, keyword, or continue" --> G[A Shape jumps in]
F -- "chat is quiet" --> H[Always have something to say: one Shape speaks]
F -- "no trigger" --> I[Shapes stay quiet]
```
The big takeaways: **@mention** always wins, **activation** makes a Shape a reliable responder, and **Free Will** governs everything else. Replies happen one after another, not all at once, so even a full chat reads cleanly. Now let's tune that behavior.
## The main dial: Free Will
**Free Will** decides when a Shape is allowed to speak on its own. It's set per chat in **Chat Settings → AI Configurations**, and it's the single most important lever for social behavior.
| Trigger | What it does | Use it when |
| -------------------------------- | ----------------------------------------- | ------------------------------------------------ |
| **Reply when mentioned** | Responds when someone uses its @name. | Always. This is the baseline. |
| **Keep the convo going** | May continue after it has replied. | You want momentum, not one-shot answers. |
| **React to keywords** | Jumps in when watched keywords appear. | The Shape owns a topic (a character, a subject). |
| **Always have something to say** | Replies when nobody else does. | A quiet chat you want kept alive. |
| **Come back later** | Checks back in after the chat goes quiet. | A companion that re-engages, not a one-off. |
The art is in the *combination*. Each trigger you add makes the Shape more forward. Stack them all and you get an extrovert that never stops; enable only "reply when mentioned" and you get a tool that waits to be called.
You can also turn on **time awareness** (so the Shape knows when it's late, or that hours have passed) and control how its messages appear when they span multiple lines.
## Recipes: tune Free Will to the chat
There's no universal setting. The right setup depends on what the chat is *for*. Start from one of these and adjust.
**Mentioned + Keep the convo going + Always have something to say + Come back later.** You want the Shape to carry energy and revive a dead chat. This is the most forward setup: great for a companion, too much for a work chat.
**Mentioned + Keep the convo going + React to keywords.** Lots of people are already talking, so let the Shape stay in character and jump in on its topic. Skip "always have something to say" so it doesn't bury everyone.
**Mentioned + (optionally) Keep the convo going.** In a [work chat](/ai-at-work), silence is a feature. The Shape should answer when called and otherwise stay out of focused discussion.
**All triggers on.** In a private chat there's no one to talk over, so maximum presence usually feels best.
## Reading people: personas and memory
Speaking at the right time is half of it. Knowing *who you're talking to* is the other half.
* **[Personas](/personas)** are how a Shape sees each person: the name to call them and what to remember. A persona is why a Shape can greet two people in the same chat completely differently. Each person sets their own.
* **Memory** gives the Shape continuity. [Short-term memory](/memory) keeps it coherent within a conversation. Long-term memory lets it recall an inside joke or a decision from days ago. A Shape that remembers feels like a participant; one that forgets feels like a vending machine.
Design these together. A Shape with good Free Will *and* good memory doesn't only reply on time. It replies like someone who's actually been in the chat.
One quiet bit of social grace is built in: when a person directly @mentions *another person* (not the Shape), Shapes hold back and let the humans talk. Your Shape isn't going to barge into a side conversation that was clearly aimed at someone else.
## Coordinating multiple Shapes
Group chats can hold **up to five active Shapes**, and they reply one after another rather than all firing on the same line. That ceiling lets you build a small cast without burying the humans. (You can also @mention a Shape to call on just that one.)
To make multiple Shapes feel like an ensemble instead of an echo:
* **Give each one a distinct role and voice.** A coordinator, a researcher, and a jokester won't step on each other the way three near-identical Shapes will.
* **Differentiate their Free Will.** Maybe one is keyword-driven and one only replies when mentioned, so they don't all fire at once.
* **Use [activation](/chats-101) deliberately.** Activate only the Shapes a given conversation needs; deactivate the rest so the chat stays legible.
## Group vs. DM: the chat changes the behavior
The same Shape should behave differently depending on the chat:
* **In a one-on-one DM**, there's no one to talk over. Maximum presence is good, so turn all of Free Will on, let it carry the conversation, and write at whatever length fits.
* **In a group chat** with several humans, restraint is the whole game. Activated Shapes reply in turn, so two chatty Shapes plus five people gets loud fast. Dial Free Will *down*, keep replies short, and let the humans drive.
The trap is designing only for the DM where you test it, then dropping the Shape into a busy chat where it won't stop talking. **Design for the crowded case.** A Shape that behaves in a group is always fine in a DM. The reverse isn't true.
## Being funny without being annoying
Personality is encouraged here (we literally say the AI is *that good*), but in a group the line between "funny" and "exhausting" is real. What separates them:
* **React, don't perform.** Respond to the actual joke someone made; don't manufacture a bit on every message. The funniest Shapes are mostly quiet and then land one line.
* **Say it once and get out.** One sharp reply beats three okay ones. Length is the enemy of comedy in a fast chat.
* **Frequency is everything.** In a busy chat, turn *off* "always have something to say." A joke every message is noise; a joke every tenth message is a personality.
* **Callbacks beat new jokes.** With [memory](/memory) on, a Shape can bring back the running bit from last week, which feels human in a way a fresh one-liner never does.
* **Never explain the joke**, and never @ the whole chat for a laugh.
## Reading the chat when humans clash
Group chats get tense or awkward, and how a Shape handles that is pure social intelligence:
* **Default to staying out.** When people are arguing, the right move is usually silence. Keep Free Will modest so the Shape isn't compelled to weigh in, and never let it pile on a side.
* **Help only if asked.** If someone turns to it and asks "what do you think?", a good Shape de-escalates or lightens the mood instead of declaring a winner. Bake that into the preset ("stays neutral in disagreements, defuses with humor").
* **Awkward silence is a judgment call.** Sometimes a quiet chat is comfortable, sometimes it's dead. "Come back later" can revive a genuinely stalled chat, but don't build a Shape that refuses to let a calm moment breathe.
* **The human stuff is the chat's job.** Slow mode, blocked words, and bans (see [Moderation](/chats-101)) handle people behaving badly. That's not your Shape's responsibility, and it shouldn't try to police it.
## Voice in a crowd: presets and personality
A Shape's [preset](/presets) and personality decide *how* it talks. In a multiplayer setting, two things matter more than in solo chat:
* **Length.** A Shape that writes essays will take over a fast group chat. Favor shorter, punchier replies. The [When Less Is More](/shortguide) guide makes the case well.
* **Reactivity over monologue.** Write the personality to *respond to people*: react, ask, build on what was just said, instead of delivering speeches. Prompt craft for this lives in [Prompt 101](/prompt101) and the [Prompting Guides](/promptingguides).
## Chat-level behavior: instructions and guardrails
Some behavior belongs to the *chat*, not the Shape. Set it once at the chat level:
* **[Custom Chat Instructions](/custominstructions-chats)** apply to every Shape in the chat: the scene, the rules, the tone, who's who. Put shared context here instead of repeating it.
* **[Initial messages and a Welcome Shape](/fandomchat)** set the first impression when someone joins.
* **[Moderation](/chats-101)** (slow mode, character limits, blocked words) keeps a busy chat from turning into chaos, human or AI.
## Test, watch, iterate
Social intelligence is tuned, not guessed. Once your Shape is in a chat:
1. **Watch real conversations.** Is it talking too much? Too little? At the wrong moments?
2. **Adjust one thing at a time**, usually a single Free Will trigger or the reply length.
3. **[Regenerate](/regeneration)** a few responses to see the range of behavior before deciding.
4. Repeat until the Shape feels like it belongs in the chat.
## The checklist
A socially intelligent Shape usually has all of this dialed in:
* Free Will matched to the chat's purpose, not maxed out by default.
* A clear role, especially if other Shapes share the chat.
* Memory and a persona so it knows the people and the history.
* Replies that are short and reactive, not long monologues.
* Chat-level context in chat instructions, not repeated every message.
* Tuning based on watching it live, not assumptions.
## What's next
Your Shape looks good, thinks well, and behaves in a crowd. The last chapter is about getting it in front of people and making it better over time.
Publish it, help people find it, and keep improving it.
[Start designing on Shapes](https://shapes.inc)