Module 2 · Working with Nomic
Working with the Nomic agent
The Nomic agent is the agent behind every part of Nomic. Drawing and BIM model reviews, submittal and RFI responses, spec and archive research, code checks, and workflows all run on it with a different brief. Learn to work with it in a conversation and the rest of the platform follows.
What the Nomic agent is made of
- Instructions. The system prompt and the built-in procedures for design and construction work. Ask for a drawing review and the survey-scope-review procedure loads based on how you phrased the request. Nothing to switch on.
- Tools. Search Files, Read File, Render Page, artifacts, BIM and CAD queries, sub-agents, and Codes Search when enabled. See Tools.
- Model and effort. Which foundation model, and how long it reasons before acting.
You set the third from the input bar. The first two respond to what you ask for.
Writing the prompt
Two ways to ask for the same review.
Review these drawings.
The agent has to guess the standard, the depth, the disciplines, and what counts as a finding, so the result may be less specific than you wanted: a long list of minor inconsistencies instead of the five things a contractor would catch.
A constrained prompt fixes that by being more specific. Pick the role closest to yours. The shape is the same in all three examples below.
Review sheets A-201 through A-210 in the attached set against the firm drawing standard in the project files (Drawing Standards v4.pdf).
Focus on: missing or wrong sheet references in callouts, detail bubbles that point to details that do not exist, and door tags with no matching schedule entry. Skip cosmetic issues.
Only flag what a general contractor would raise as an RFI.
Compare the structural notes on S-001 and S-002 with the foundation plans S-101 through S-104 in the attached set.
Focus on: footing or slab thicknesses and reinforcement that differ across the notes, schedules, and plans, and also details called out in a schedule that are not in the set. Skip cosmetic issues.
Compare the door hardware sets in spec section 08 71 00 with the door schedule on A-601 in the attached set.
List every door whose hardware set does not match, with the door number, both values, and the page of each. Skip doors that match.
Put results in a table the superintendent can work from.
Each prompt names the files and bounds the scope. Specify the output only when you need something particular, like the table the superintendent can work from in the last example. The standard review output, a findings table with citations and severity, comes with every review.
A lookup prompt does not need to be this constrained. "Which spec section covers roof membrane warranties?" is fine as written. A review needs a constrained prompt because it is only as good as the brief.
- Name the files. "The attached set" is ambiguous with four PDFs attached. "A-201 through A-210" is clear.
- Define a finding. "What a contractor would raise as an RFI." "Skip cosmetic issues." "Structural against architectural only." "Anything that changes scope or cost."
- Name the output when you need something particular. A table, a memo, an annotated PDF, a one-line answer.
Where to ask from
For a question about a project, ask from inside the project. The agent has the project's files as standing context and the conversation appears on the project page, where teammates can view it once you share it with the project.
For a one-off question about a file outside any project, start a session and attach the file with +. The file is available for that conversation only.
Managing a session
The conversation accumulates context: messages, replies, and everything the agent read. The conversation's context has limits.
- One task per session. Switching from the drawing review to an unrelated RFI: new conversation.
- Continue when the context helps. Still on the same review: follow-ups are cheap and good. "Go deeper on the door and finish schedules." "Only the Level 2 findings."
- Start fresh when it drifts. Repeating itself, losing track of which sheet: a new session beats a fifth clarification.
- Effort can change mid-session. The model cannot. Raise effort when a follow-up needs more of the set read. To change the model, start a new session.
Past sessions are in the sidebar and on the project page. Search chats finds them. Session owners set visibility (Private, Shared in Project, or Shared in Org on Business and Enterprise) and can Move to Project if a conversation started in the wrong place. Shared conversations are read-only for recipients.
Common failure: asking with nothing attached
A project question in a session with no project and no attachments gets an answer from general knowledge, fluent but not about your building. The agent can describe what a typical slab edge detail looks like, but it cannot say what yours calls out.
If an answer reads like a textbook, check what is in scope before rewording. See Context in the agent guide.
Quiz
You want the agent to flag missing detail references on ten sheets and nothing else. Which prompt?
Next: setting up a project so the agent can find what it needs.
Next: Understanding your project