Instructions
A Project's Instructions are plain English guidance the agent reads at the start of every conversation and every work item. This is where tone, scope, and the rules your team already lives by get written down. You are not coding the agent; you are explaining how to show up.
Audience: builders
Write these as you learn them, not before. A brand new Project ships with a template, and it is tempting to fill it in up front. It is usually better to start working and come here when the agent gets something wrong in a way you can name. An instruction written to fix a real mistake is specific; one written in advance tends to be a description of your team that changes nothing.
Where it lives
Open your Project in Build. Instructions is its own section in the Context tree, below Files. The editor opens with whatever instructions already exist.
The editor saves on a debounce. Watch the last-saved timestamp ("just now", "5m ago") to confirm your changes landed. The file itself is Instructions.md in the Project's volume, so it also shows up in the file browser if you want to edit it that way.
Start from the template
Every new Project ships with a starter Instructions.md. Replace the example copy with your team's reality, but keep the structure. It works because each section tells the agent a different thing:
- Purpose. What this Project is for, in one or two sentences, in your team's language.
- Voice and tone. How the agent should write. Be specific. "First person plural, active voice, no buzzwords" beats "professional".
- What to focus on. The work the agent should help with most often.
- What to avoid. Things the agent should not do. Name the patterns you have seen go wrong.
- Sources of truth. Where the agent should pull context from. Point at specific files, integrations, and Data Collections by name.
- How to ask for help. When the agent should ask before guessing instead of inventing an answer.
Drop sections you do not need. The headings are a starting outline, not a contract.
Watch the word count
The editor surfaces two soft limits as banners:
- Around 2,000 words: yellow banner. Past this point your instructions start competing with the per-turn context budget. Move reference content into separate files and point at them from Sources of truth.
- Around 4,000 words: red banner. The prompt layer truncates the file beyond this length, so the agent stops seeing the tail. Trim or split.
Heavily-used Projects do better well under 2,000 words. Style guides, vendor lists, and glossaries belong in separate files the instructions point at, not in the instructions themselves.
Iterate as you go
There is no publish step for instructions. Save, and the next conversation turn or work item picks up the latest version. Anything already running keeps the version it started with.
That means you can edit instructions while a conversation is open, then send another message to see the change take effect. Use this loop to refine voice, expand "What to avoid", or tighten "Sources of truth" without losing the thread.
The same edit reaches your Workflows without any further step, so a correction you make once applies both to the people chatting with the Project and to the work it runs on its own.