Building workflows

The Nomic agent knows how construction documents are organized, how a submittal review is structured, and what a contractor raises as an RFI. It does not know how your team does it: which checklist, which severity scale, which columns the PM wants, and which sheets to skip. You can say so in every prompt, or write it down once as a workflow.

When a prompt becomes a workflow​

Start in a conversation. When a prompt is earning its keep, package it. From Workflows, reach for a workflow when:

• The task repeats. The tenth submittal review should not depend on how the tenth person phrased it.
• Output must be comparable. Same sections, columns, and severities across people and projects.
• Reference material is permanent. The checklist or standard stays attached instead of being re-uploaded.
• Runs are long or unattended. An Extra High run can take a while, and workflows can be launched from the Agent API.
• Someone else runs it. On Business and Enterprise, reviewers fill a file-request form instead of writing a prompt.

A workflow does nothing the agent cannot do in a conversation. A workflow makes the good version the default.

What a workflow is​

Part	What it is
Instructions	The brief: what to do, in what order, in what format.
Tools	Which capabilities the agent may use, one row each: Codes Search, Web search, and MCPs. All are off until you add them, including on cloned Nomic workflows.
Context	Files and folders available on every run. Checklists, standards, templates. You add these yourself; the builder agent cannot.
Requested files	Per-run inputs the user attaches: this week's set, this submittal. Slots can be required or optional, single or multiple, and can accept folders.
Project scope	Optionally tie the workflow to a project so it lives on that project's Workflows page.
Model and effort	The model, and Low, Medium, High, or Extra High, fixed for every run. You choose these; the builder agent does not.

Context versus requested files is the distinction that matters. Context is what is always true: the checklist. Requested files are what changes each run: the submittal being checked. Get that split right and one workflow serves every project.

Open figure

Start from a Nomic workflow​

Workflows → Nomic lists starter workflows Nomic ships and maintains. From Nomic workflows:

• Drawing: Drawing Standards Creation, Drawing Standards Review, Residential Drawing Standards Check
• Coordination: Submittal Review, RFI Response
• Preconstruction: RFI Anticipator, Submittal Log Generator, Drawing Revision Analysis
• Operations: Invoice Processing Automation, RFP Response Drafter

Run one as-is to see the output. To match your team, Clone it. The copy opens in the builder with the builder agent ready.

The starter that pays off first depends on the work you repeat:

The example through the rest of this lesson is Submittal Review, because every seat touches submittals.

Check submittals against our review checklist as well as the spec. The checklist is attached as context. Tighten the instructions so every finding cites the submittal page and the spec paragraph or checklist item, and make the submittal and the specification section required file slots.

Instructions​

Instructions decide whether a workflow holds up. Authoring instructions recommends five sections, in order:

1. Role: what the agent is doing and for whom. "A thorough review of a submittal against the attached spec section and review checklist, for the reviewer of record." Not a persona.
2. Output format: exact table columns or report sections. For a submittal review: Submittal Page, Spec Paragraph, Deviation, Recommended Action.
3. Numbered steps: survey, per-section checks, assembly. Use the Explore sub-agent for reconnaissance before the expensive work.
4. Tool instructions: when to dispatch sub-agents (one per spec section or issue class, in parallel), when to render a page rather than read text, how to record findings.
5. Verification: every finding cited, no duplicates, severity assigned.

Two rules from the same guide: put the logic in Instructions, not in linked artifacts, which are reference data only. Use sub-agents deliberately; each reads its own slice of the set.

Sharing and releasing​

New workflows start Invited people only. Add collaborators from the Sharing tab, and on Business and Enterprise choose organization-wide or project scope. Edit in a draft, test with Run draft, release a version when it is right. Teammates run the released version while you iterate. See Workflow management.

The spend estimate from lesson 3 shows on the workflow row and the run form and sharpens with runs.

Open figure

A full build, start to finish, from the October 7 Office Hours:

Before and after​

Before: five reviewers, five prompts, five differently shaped tables, one against last year's checklist.

After: one released workflow with the current checklist in Context, required Submittal and Specifications slots, fixed columns, Medium effort. Every reviewer gets the same table. When the checklist changes, one file changes.

Common failure: the QA manual in Instructions​

Instructions are context on every run. A very long brief costs tokens every time and makes the agent worse at following any one part of it. Keep role, format, steps, and verification tight. Put the checklist itself in Context, where the agent reads the parts it needs.

The mirror failure is an output section that says "a table." Name the columns, or each run produces a different one.

Next: everything on one project, start to finish.

Next: Putting it all together