> ## Documentation Index
> Fetch the complete documentation index at: https://docs.concord.ad/llms.txt
> Use this file to discover all available pages before exploring further.

# Best practices

> How a conversation works in Concord, and how to get reliable results from it.

Concord works best when you know what it sees, what it keeps, and what it asks before acting. This page explains it in practical terms.

## What Concord knows when you send a message

At every message, Concord works with:

* Your **organization instructions** and your **personal instructions**.
* The summaries of your workspace's [Skills](/features/skills/understanding-skills). Concord reads them, picks the ones that apply to your request, and loads them in full before acting.
* The platforms and accounts you have access to.
* The files in the conversation, the platform actions waiting for your review, and the current plan.
* Everything said in the conversation so far.

**Name what you work on.** Give the advertiser and the campaign by name, ID, or link in your first message, so Concord does not have to guess.

**Name a Skill when it must apply.** Concord selects Skills from their summary. If a rule must apply to this request, say which Skill to follow.

## What carries from one conversation to the next

A conversation does not remember another. What you agreed in one thread is not known in a new one.

What applies to every conversation:

* Organization and personal instructions
* Skills
* Your connected accounts and your workspace's scope

**Turn recurring rules into Skills.** If you find yourself repeating a naming convention, a budget split, or a QA rule, save it as a [Skill](/features/skills/creating-skills) or add it to your [instructions](/settings/organization-description). Concord may offer to save a rule as a Skill when you describe one.

## How Concord keeps track of a long task

For a request with several steps, Concord writes a **plan** first and follows it step by step. The plan stays in view for the whole run, and you can follow progress in it.

Alongside the conversation, Concord keeps a running record of what the conversation has established:

* what it created, updated, or failed to change on your platforms, and the files it produced;
* the rules and decisions you gave it ("no more than €5 CPM", "keep the existing naming");
* your answers to its questions;
* what is still open.

It is a record of facts, not a transcript: when a decision changes, the new one replaces the old. Concord reads it at every message, so what you settled early in the conversation still applies an hour later. It never stands in for the live state of your platforms, which Concord reads again when it needs it.

When a thread gets long, Concord also condenses earlier messages: it keeps the recent exchange, the platform identifiers it resolved, and the files you shared, and summarizes the rest. See [Long conversations](/features/chat/long-conversations).

**State constraints and decisions explicitly.** "Never above €5 CPM", "exclude the brand safety list", "flight ends 30 June": a clear sentence is recorded as a rule. A hint in passing may not be. When you change your mind, say so plainly, so the new decision replaces the old one.

**Give multi-step jobs in one message.** Concord plans the whole job at once and does not lose a step between messages.

**One task, one conversation.** Start a new conversation for a new campaign or a new question. It starts clean, and you restate only what matters.

**Trust the review, not the prose.** The [platform actions](/features/platform-actions/platform-actions-overview) waiting for review are the record of what Concord prepared. Check them before you submit.

## While Concord is working

You do not have to wait. Type a correction or an extra detail and send it: Concord reads it at the next step without losing what it was doing. Click **Stop** if it is heading in the wrong direction.

## What Concord asks before acting

Reading data, running reports, and preparing platform actions never change anything on your platforms. Preparing only adds a platform action to your review.

Submitting platform actions follows your **Approvals** setting in [Profile settings](/settings/profile-settings):

* **Always ask**: you click to confirm every submission.
* **Auto**: Concord submits on its own only when your last message is an explicit go-ahead for exactly that change. "OK" or "sounds good", a question, or a change of values brings back the confirmation buttons.

**Say exactly what you approve.** In **Auto**, "Yes, submit the three budget changes" goes through. "Looks fine, but maybe lower the second one" does not, and Concord asks again.

## How fresh the numbers are

* **Reports** are fetched from the platform when you ask for them.
* **Lists of campaigns and entities** come from Concord's copy of your accounts, refreshed as you work and every night for running campaigns.
* **Today's figures** are partial until the day closes on the platform. Concord flags them.

**Ask for a report when a number must be exact.** If you just changed something directly in the platform, give Concord the ID or link of the object, or ask it to refresh.

## Working with files

* Attach a file and say what to do with it in the same message. A file sent alone gets a short description and a question back.
* Files up to 500 MB are accepted. Spreadsheets (CSV, XLSX), PDFs, images, Word, and PowerPoint documents are read.
* Targeting lists in a file are detected and saved so you can reuse them.
* After a long pause or once a conversation has been condensed, Concord reads a file again before quoting figures from it.

See [Files](/features/files/files-overview).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.