Identity, skills and memory

Use Instructions to maintain behavior. Keep identity short, put recurring procedures in skills, and use memory for validated lessons.

Before you begin

You need owner or dashboard editor access to edit instructions. Prepare representative input files and expected outputs for document-based tasks.

Identity

1. Open the agent's Instructions tab.
2. Select Edit on Identity, or Write identity if an older agent has none.
3. Describe its role, scope, decisions, hard rules, tone and when to ask a human.
4. Select Save.

Identity lives in AGENTS.md and accompanies every conversation and run. An agent without identity cannot receive work. Move procedural detail into a skill so each run carries only what it needs.

Identity is formatted text: type / to add headings, lists, tables or images, and @ to link a file the agent can open. Images you add are saved in the agent's Home, next to AGENTS.md. Choose Source to edit the Markdown itself.

Skills

1. Select Add skill, then Write a new skill.
2. Give the skill a name and a description that states both the task and when to use it. For example: “Prepare a quote when a client requests a proposal or custom pricing.”
3. Write the procedure as numbered steps, rules and examples. Reference required sources and define a checkable result. Type / for headings, lists, tables or images; once the skill exists, images are saved in its folder.
4. Add templates, examples or price lists with Add files or Add a folder.
5. Type @ to insert a Markdown link to a support file or Workspace file. Shared files still require the agent's access.
6. Select Create skill, or Save when editing.

At run start, the agent sees skill names and descriptions. A matching description leads it to load SKILL.md, then referenced files. Scripts run through Code Executor.

Skills use the open Agent Skills format: skills/<name>/SKILL.md contains name and description in its header, followed by Markdown instructions. References, assets and scripts can use relative paths inside the skill folder.

To rename a skill, change its name on the skill page and save: its folder and support files follow. Workspace never renames or moves a skill folder or its SKILL.md. Other documents that mention the old name keep it.

Added and removed support files apply when you save; removed files go to Trash. If an upload fails, keep the draft and retry. If someone saved a newer version, compare and merge it before saving; your draft does not silently overwrite it.

Deleting a skill moves its whole folder to Trash. You can restore it for seven days.

Import a skill

1. Select Add skill, then Import from a command or URL.
2. Paste a repository URL or an npx skills add command.
3. Select Preview.
4. Review the selected skills, files, scripts, compatibility and audit status.
5. Select Import to install the reviewed copy.

Botchi parses the command as text; it never executes it. You can also use Browse the community catalog and Import on a skill's detail page. Browsing installs nothing.

An installed skill is an account-owned copy. Imported identifies its source; Modified from source means a file changed locally. Runs do not depend on the source remaining online. Failed security audits and name collisions block installation; source changes require another preview.

Scripts

1. Open Workspace on the agent.
2. Store reusable Node.js or Python code in scripts/, or bundle it with a skill.
3. Reference the script and required input files in the skill that uses it.
4. Test the result in Chat, then rerun evals.

Use skills for procedures requiring judgment. Use scripts for repeatable calculations, conversions and data processing. Code Executor reads and writes only within the agent's Workspace access.

Memory

1. Open Instructions and find Memory.
2. Select Open to review MEMORY.md in the file editor.
3. Correct outdated mappings or lessons and remove duplicated company facts.
4. Save your changes and test a task that depends on them.

The agent curates memory and reads it when past context matters. Keep durable workflows, verified mappings, tool quirks and corrections. Consolidate rather than append indefinitely.

Related

• Test behavior with evals
• Give an agent tools
• Workspace access