# Instructions

> Standing context about who you are and how you want to be answered, applied automatically to every prompt.

Source: https://triplo.ai/docs/instructions  
Last updated: 2026-10-18

## What an Instruction is

An Instruction is context you would otherwise retype constantly: who you are, what you work on, and how you want to be spoken to. Select one and it is attached silently to **every** prompt until you change it.

Without one, you find yourself writing "I'm a backend engineer, be concise, don't explain basics" at the top of half your prompts. With one, you write it once.

### Two halves

- **What**: Who you are and what you do. Role, industry, seniority, the systems you work with, your audience. Up to 5,000 characters.

  Answers the question *"who am I talking to?"*
- **How**: The shape of the reply. Length, tone, formatting, what to leave out. Up to 2,000 characters.

  Answers the question *"how should I say it?"*

Both are sent as one system message, with the *How* text prefixed by `Response style:`. Keeping them separate is what makes an Instruction reusable — you can rewrite how you want answers without touching who you are.

## Creating one (Paid plan)

**How to reach it.** Action grid → Instructions → Add Instruction….

1. **Title it for the situation, not for yourself**: *Client-facing email* beats *My instruction 2*. You will be picking from this list several times a day.
2. **Fill in What**: Concrete beats abstract. "I maintain a Django monolith and a small React admin panel" is worth more than "I'm a developer".
3. **Fill in How**: State what you do *not* want as well as what you do. Models over-explain by default, and this is where you stop them.
4. **Tag it and save**: Tags make the drawer searchable once you have a dozen.
5. **Select it**: Creating an Instruction does not activate it. Choose it from the book icon in the top bar.

> **Warning:** Both **What** and **How** are required. An Instruction with either left blank will not save.

### Active versus selected

Two different switches, easily confused:

- **Active** — the toggle in the drawer. An inactive Instruction is hidden from the selector. Use it to retire something without deleting it.
- **Selected** — chosen in the top bar. Only one at a time, and only a selected Instruction affects your prompts.

## Worked examples

### Backend engineer

```
What:
Senior backend engineer. Python and Go. Postgres, Redis,
Kubernetes on GCP. I work on a payments service where
correctness and idempotency matter more than elegance.
Comfortable with distributed systems; no need to explain
CAP, eventual consistency or basic SQL.

How:
Lead with the answer, then the reasoning. Code first,
prose second. Point out race conditions, partial failures
and rollback paths without being asked. Never explain what
the code does line by line. If my premise is wrong, say so
in the first sentence.
```

What that buys you is easiest to see side by side. Below, the same question goes to the same model twice — once with nothing selected, once with the Instruction above. Nothing else changes: no Mind, no SmartPrompt, no extra words in the prompt.

*Animation: The first answer is not wrong. It is just written for someone who has never seen a lock before — and you read four paragraphs to find that out.*

### Client-facing email

```
What:
Account manager at a mid-sized design agency. My clients
are marketing directors at consumer brands. They are busy,
non-technical, and often reading on a phone.

How:
Warm but efficient. Under 150 words. Short paragraphs, no
bullet lists. Never use "just", "simply" or "reach out".
Always end with one clear next step and who owns it.
No greeting or sign-off — I add my own.
```

### Contract review

```
What:
In-house counsel at a SaaS company, England and Wales.
I review supplier and customer contracts. I need risk
flagged, not summarised.

How:
Answer as an ordered list of concerns, most material first.
Quote the exact clause text you are referring to. State the
risk and a suggested redraft. Say plainly when a clause is
standard and needs no action. You are not my lawyer — do
not caveat every paragraph, one note at the end is enough.
```

### Support agent

```
What:
Tier-2 support for a project management tool. Customers
reach me after tier 1 has failed, so they are usually
frustrated and have already tried the obvious.

How:
Acknowledge the problem in one sentence — no apology
paragraphs. Then numbered steps, one action each. Assume
they can find settings but not developer tools. If I need
more information, ask at most two questions and say why
each one matters.
```

## Getting more out of them

### One per mode of work, not one per project

Three or four Instructions covering how you work — *writing code*, *writing to clients*, *reviewing*, *thinking out loud* — will serve you better than fifteen tied to individual projects. Project-specific knowledge belongs in a [Mind](https://triplo.ai/docs/training), which is retrieved on demand rather than carried on every prompt.

### Say what to omit

The highest-value lines in a *How* are usually negative. "No preamble." "Don't restate my question." "Don't hedge." Each one removes something you would otherwise skim past dozens of times a day.

### Pin one to a prompt with a Magnet

If a SmartPrompt only makes sense under a particular Instruction, attach it as a [Magnet](https://triplo.ai/docs/magnets). Triplo switches to it for that prompt and switches back afterwards — no manual toggling.

### Instruction or SmartPrompt?

|   | Instruction | SmartPrompt |
|---|---|---|
| **Applies to** | Every prompt, until changed | Only when you invoke it |
| **Describes** | You and your preferences | A specific task |
| **Good for** | "I'm a nurse; use clinical terms" | "Turn these notes into a handover" |
| **How many active** | One | As many as you like, in sequence |

Rule of thumb: if it is true of you regardless of the task, it is an Instruction. If it describes a task, it is a SmartPrompt.

## Limits & gotchas

- **Title** 100 characters, **What** 5,000, **How** 2,000. Both text fields are mandatory.
- The Instruction is sent on every prompt, so it costs tokens on every prompt. A 5,000-character *What* is roughly 1,250 tokens each time.
- It applies to background hotkey prompts too. An Instruction that says "always ask a clarifying question first" will ruin a prompt whose output gets pasted straight into your document.
- A conflicting SmartPrompt is not overridden by the Instruction, nor the reverse. Both are text; the model decides. Keep them compatible.
- Instructions are shared to your other devices as **read-only** copies. Clone one to edit it locally. See [Sharing across devices](https://triplo.ai/docs/sync#shares).
- Each Instruction is a Markdown file under `~/Documents/Triplo AI/Instructions/`, editable directly — see [File formats](https://triplo.ai/docs/advanced#file-formats). Files written by older versions using `Context` and `Response Style` headings are still read correctly.