Add context, files, and instructions to a workflow

Every workflow combines three layers of knowledge:

Instructions — written guidance for how to do the task.
Context files — documents always available on every run (standards, specs, prior approvals).
Requested files — slots the user fills in when they run the workflow (today's drawing, this submittal, this RFI).

Use the right layer for the right kind of content and your workflows stay accurate as projects change.

Instructions

Instructions describe the procedure: what to check, what tone to use, what to skip, what to flag.

Write them like you'd write a short SOP for a teammate, not as a prompt to an AI.
Spell out edge cases. "If the sheet has no scale bar, note it and continue."
Include the output format you want — bullet list, table, marked-up PDF — and Nomic will follow it.
Include a final verification step when the task needs every page, section, or item checked.

Use context files for things every run should consider: project standards, code references, design narrative, glossary of trades.

Avoid putting one-off inputs in context — they bias every run.

Use requested files for the inputs that change every run: today's drawing set, the specific submittal, this week's RFI log.

Define a named slot in the builder (e.g. "Drawing set", "Submittal package").
When a user runs the workflow, they upload or pick a file for each slot.
The slot name is what the user sees in Assistant, so make it descriptive.

Putting the input in context instead of a requested file — every run will analyze the wrong drawing.
Vague instructions — "check this drawing" gives unpredictable results; "flag any duct shown crossing through a structural beam" is reliable.
Too much context at once — pruning to the standards that actually apply for this task improves precision.
Skipping a sample run — broad workflows are harder to debug after a full drawing set has already run.