• Customers
  • Docs
  • Pricing
Log inLGet startedS
Instructions and skills

Platform

Start here

  • Get started
  • Use the API
  • CLI reference

Concepts

  • Sandboxes
  • Commands
  • Model providers
  • Files
  • Images and registries
  • Volumes
  • Secrets
  • Endpoints
  • Connections

Reference

  • SSO and SCIM
  • Organizations and access
  • API reference

Company agent

Start here

  • Get started
  • Core concepts

Set up

  • Models
  • Sandboxes
  • Knowledge
  • Integrations
  • Secrets
  • Instructions and skills
  • Members and roles

Use Adapt

  • Agents
  • Chats and tool activity
  • Automations

Surfaces

  • Slack
  • GitHub
  • Email
  • iMessage and SMS
  • Webhooks

Run Adapt

  • Self-host
  • Security and privacy

Platform

Start here

  • Get started
  • Use the API
  • CLI reference

Concepts

  • Sandboxes
  • Commands
  • Model providers
  • Files
  • Images and registries
  • Volumes
  • Secrets
  • Endpoints
  • Connections

Reference

  • SSO and SCIM
  • Organizations and access
  • API reference

Company agent

Start here

  • Get started
  • Core concepts

Set up

  • Models
  • Sandboxes
  • Knowledge
  • Integrations
  • Secrets
  • Instructions and skills
  • Members and roles

Use Adapt

  • Agents
  • Chats and tool activity
  • Automations

Surfaces

  • Slack
  • GitHub
  • Email
  • iMessage and SMS
  • Webhooks

Run Adapt

  • Self-host
  • Security and privacy

The open platform for company agents. Use the whole platform or build with individual components.

Products

Company agentAI Transformation ServicesSandboxesStorageModel gatewayPricing

Solutions

EnterpriseFor CEOsEngineeringOperationsSalesAll solutions

Resources

DocumentationDeveloper guideChangelogBlogCustomer storiesIntegrations

Company

AboutCareersTrust centerSupportTalk to usGet started

Social

XLinkedInGitHubYouTube
© 2026 AdaptMade with ♥ in Adapt teammates walking outside the Presidio officeAn Adapt engineer at their desk at dusk, the Golden Gate through the windowAn Adapt teammate pointing to the Adapt sign at the office doorTermsPrivacy

Set up

Instructions and skills

On this page
  1. Write company instructions
  2. Set Agent defaults
  3. Add skills
  4. How agents use skills
  5. Manage skills

Company instructions are guidance every agent in a workspace follows; skills are reusable procedures agents load when a task calls for them. Agent defaults set the starting point for new agents.

Owners and admins change all three. Members can read them, and set instructions and settings on the agents they own.

Write company instructions

Edit them at the top of Settings → Defaults, under Company instructions. They apply to every agent on every run, including agents that already exist, so a change takes effect on the next message.

Each run's instructions are assembled in this order:

  1. Company instructions.
  2. The agent's own instructions.
  3. The list of available skills, when the agent has harness tools.

Use company instructions for what holds everywhere: how to cite sources, what never leaves the company, tone, who to ask. Put role-specific guidance in each agent's instructions.

  • Length: up to 32,000 characters.
  • A new workspace starts with: "You are a helpful company agent. Be precise, cite your sources, and protect company information."

Set Agent defaults

Settings → Defaults holds the settings new agents start with. Saving them doesn't change existing agents: each agent copies the defaults once, when it's created, and keeps its own settings after that. Company instructions and the Off for everyone setting below are the exceptions; they apply to every agent at once.

Capabilities

  • Models: the model connection and exact model ID new agents use, plus optional image and decision models. See Models.
  • Knowledge, Sandbox and Integrations: the connections new agents start with. See Knowledge, Sandboxes and Integrations. Defaults can't select secrets; choose those on each agent.
  • Harness tools: plan ahead, load skills, search past chats and access workspace documents.
  • Manage Adapt: set up agents, providers and skills from chat.

Knowledge, Sandbox, Harness tools and Manage Adapt each have an availability setting:

  • Off for everyone: no agent can use it, including existing agents.
  • Off by default: new agents start without it; it can be turned on per agent.
  • On by default: copied into new agents. Existing agents keep their settings.

Limits

Limits are copied into new agents. Leave a field blank for no limit.

  • Steps per run: model steps in one run. Default: no limit.
  • Time per run: 1 to 35,791 minutes. Default: 5 minutes. A run that hits the limit stops; runs can always be stopped by hand.
  • Output per step: at least 256 tokens. Default: 4,096. Blank uses the model's own output limit.
  • Context budget: at least 12,000 characters. Default: 60,000. Model limits still apply.

With no step limit and no time limit, a long task can delay other conversations until someone stops it.

Add skills

A skill is a folder with a SKILL.md file and any supporting files, in the Agent Skills format. Skills belong to the workspace: every agent with harness tools sees the list of enabled skills and loads one by name when it's relevant.

Open Skills from the sidebar and choose Create:

  • Create with an agent: opens a chat where you show the agent the task once and it writes the skill.
  • Import from GitHub: enter a repository as owner/name or paste its link, pick a branch, and optionally a directory. Adapt lists the skills it finds, and you choose which to add. Public repositories only.
  • Upload a skill: a single SKILL.md file.

The Skills page also suggests skills from a catalog of public repositories, which add the same way as an import.

SKILL.md starts with YAML frontmatter:

---
name: weekly-report
description: Use when asked for the weekly status report. Pulls last week's closed issues and drafts the summary in the company template.
---

1. List issues closed in the last seven days.
2. Group them by project.
3. Fill in templates/weekly.md and share the draft.

name and description are required. Adapt also reads license, compatibility and metadata; other fields, allowed-tools included, stay in the file but aren't used.

  • Name: lowercase letters, digits and hyphens, up to 64 characters, with no leading, trailing or double hyphen. Names are unique in a workspace.
  • Description: up to 1,024 characters. Write it as when to use the skill; agents choose from it.
  • compatibility: up to 500 characters.
  • Files in one skill: up to 32 MiB together; paths up to 255 bytes.
  • Skills per workspace: 200.

Agents receive the skill list within a 20,000-character budget. When the full list doesn't fit, long descriptions are shortened in it, and the Skills page says so.

How agents use skills

The agent sees each enabled skill's name and description. When it loads one, it gets the instructions and the list of bundled files, and reads files as it needs them. If the conversation has a sandbox, the files are also written to a directory there so scripts can run.

To point an agent at a skill yourself, choose Use skill from the skill's menu. It opens a new chat with a prompt that names the skill.

Owners and admins can also ask an agent to save a skill from a chat, for example after it finishes a task you want repeated. The agent saves it for the whole workspace. The save runs with your permissions, so a member's request is refused, and agents can't save skills from surface threads or delegated tasks.

Manage skills

From a skill's menu:

  • Disable keeps the skill's files but takes it out of the agents' list. Agents can't load it until it's enabled again.
  • Sync from GitHub brings an imported skill in line with its repository's branch. Adapt doesn't sync on its own.
  • Remove deletes the skill.

Disabling a repository source disables all of its skills. Removing a source can remove the skills added from it too, if you choose.

PreviousSecretsNextMembers and roles

On this page

  1. Write company instructions
  2. Set Agent defaults
  3. Add skills
  4. How agents use skills
  5. Manage skills