# Suggestion pills

Tappable answers offered when your agent asks a question. Deterministic, instant, and live without a publish.

Source: https://docs.omazy.ai/how-to/agent/suggestions/

When your agent asks a question, it can offer the likely answers as tappable
chips instead of leaving the customer to type. You author these in the console
under **Agent, Answer suggestions**.

They are small and they punch above their weight. A customer on a phone, halfway
through a checkout, deciding whether to type "anniversary" or just close the tab,
is a real fork in the road. A chip is one tap.

## How a rule works

A rule has two halves.

| Half | What it is |
|---|---|
| **Match** | Words to look for in the question the agent just asked |
| **Options** | The chips to offer when a match is found |

If any match word appears in the agent's question, the widget shows the options.

```
Match:    Birthday, Anniversary
Question: "Welcome to Mansion Key. What are you celebrating?"
Result:   the celebration chips appear
```

The console has a **Test question** box, so you can paste a real question and see
whether your rule fires before any customer does.

## Matching is deterministic, and that matters

No model runs. It is a plain scan of the agent's question against your match
words.

Three consequences, all good:

- **It is instant.** No extra call, no added latency on the reply.
- **It is predictable.** The same question always produces the same chips. You
  can test it once and trust it.
- **It costs nothing.** Suggestions do not consume context on every turn the way
  the [brief](/how-to/agent/brief/) does.

The trade is that it is literal. It matches the words you wrote, not the meaning
around them. If your agent sometimes asks "what is the occasion" and sometimes
"what are you celebrating", both need to be in the match list.

## They go live immediately

This is the difference people trip over.

The brief needs a publish. **Suggestions do not.** Save them and they are live on
the next reply.

That makes them the fastest thing to fix in the whole agent, and also the easiest
to change without meaning to. The change is audited like any other agent edit, so
if chips appear that nobody remembers adding, the audit log knows.

## Limits

| Limit | Value | Why |
|---|---|---|
| Rules | 40 | The matcher scans every rule on every reply |
| Options per rule | 6 | More chips than that is a menu, and a menu is a wall |

The rule ceiling is a nudge as much as a limit. Past a few dozen rules you have
stopped writing suggestions and started building a decision tree, and a decision
tree is what [journeys](/how-to/engage/) are for. Journeys can branch, collect, validate
and remember. Suggestion pills can offer six words.

## Writing ones that get tapped

**Match on your agent's words, not your customer's.** The scan runs against the
question the agent asked. This is the mistake almost everyone makes first.

**Offer answers, not topics.** "Anniversary" is an answer. "Occasions" is a
category, and tapping it just asks the customer to think again.

**Keep them short enough to read at a glance.** These render as chips on a phone.
Two or three words each.

**Cover the honest option.** If "not sure yet" is a real answer to your question,
offer it. A customer who cannot see their answer types nothing at all, and a
question with no visible way out is worse than no chips.

**Do not use them to hide a bad question.** If you need six chips to make a
question answerable, the question is doing too much. Ask two smaller ones.

## When nothing appears

Work down this list.

1. **The match words are not in the agent's question.** The most common cause by
   a distance. Paste the exact question into **Test question**.
2. **The agent did not ask a question.** Chips attach to a question. A statement
   gets none.
3. **The wording drifted.** Your agent rephrases. Yesterday's phrasing is not
   guaranteed to be today's, so put every phrasing you use in the match list.
4. **Nothing was saved.** The console confirms when suggestions save.

## What this is not

Two neighbouring features get confused with this one.

**Widget starter prompts** are what a visitor sees *before* saying anything, on
the widget's opening screen. Suggestions are what they see *after* the agent
asks something. See [Widget](/how-to/widget/).

**Quick replies from a journey** come from a configured flow that collects
answers in sequence. Suggestions are a single-step convenience with no memory
and no branching.