Skip to main content
Back to Help

Improve your workflows over time

A workflow that works on the first try is rare. Most workflows need a few rounds of editing before they consistently produce output your team trusts. This page covers how to evaluate a workflow's performance and make targeted improvements.

Ask follow-up questions during a run​

When a workflow finishes its first response, you are not limited to accepting or discarding the result. You can send follow-up messages in the same conversation to refine the output before starting over:

  • Ask the agent to expand on a specific finding.
  • Tell it something was wrong and ask it to redo that section.
  • Request a different output format (e.g., "reformat the findings as a table").
  • Ask it to check an additional page or standard it missed.
  • Clarify what you meant if the output suggests it misunderstood your intent.

Follow-ups are useful for exploring what the workflow can do before you commit changes to the workflow itself. If you find yourself giving the same follow-up instruction on every run, that's a sign the instruction belongs in the workflow permanently.

Edit the workflow in the builder​

When you know what needs to change, open the Workflow Builder to make it permanent:

  1. Go to Workflows in the left nav.
  2. Find the workflow and click Edit.
  3. Update the instructions, context files, or requested file slots based on what you learned from your runs and follow-ups.
  4. Save and re-run on the same input to confirm the change helped.

The builder is where improvements become durable. Commentary you give during a run — "also check for missing fire ratings", "ignore placeholder sheets" — should eventually be written into the workflow's instructions so every future run benefits automatically.

Review outputs critically​

After each run, ask:

  • Did the output answer the right question?
  • Were findings accurate when you spot-checked them against the source?
  • Did it miss anything a human reviewer would have caught?
  • Was the output formatted in a way that's easy to hand off or act on?
  • Did it hallucinate references, standards, or drawing details?

Mark specific failures. A vague sense of "not quite right" is hard to fix — a concrete example like "missed the fire rating on sheet A2.03" gives you something to address.

Tighten instructions​

Most quality issues trace back to instructions that are too broad, too vague, or missing a step.

  • Be specific about scope. Replace "review the drawings" with "check sheets M1.01 through M1.04 for duct routing conflicts with structural members."
  • Name the standard. If the workflow should check against ASHRAE 90.1 or a project-specific spec, say so explicitly and attach the reference as a context file.
  • Describe what to skip. Tell the workflow what not to flag — otherwise it may produce noise that wastes reviewer time.
  • Spell out the output format. "Return a table with columns: sheet number, finding, severity, code reference" is better than "list the issues."
  • Add verification steps. Ask the workflow to confirm it checked every page, section, or line item before finishing.

Improve context files​

If the workflow is missing things or citing the wrong standard:

  • Add the reference it should have used (the code section, the project specification, the client standard).
  • Remove references that don't apply — extra context dilutes attention and increases cost.
  • Update stale context when standards are revised or project requirements change.
  • Split broad references into focused ones. Attaching an entire 200-page spec when only one chapter applies makes the workflow work harder and produce less precise results.

Narrow the requested files​

Workflows perform best when inputs are focused:

  • Ask for a specific sheet range rather than the entire drawing set.
  • Use separate workflow runs for different disciplines or scopes instead of combining everything into one pass.
  • Label file slots clearly so users don't accidentally upload the wrong document.

Compare runs side-by-side​

When you edit a workflow, run the new version on the same input you used before:

  1. Save the output from the original version.
  2. Edit the workflow.
  3. Run again with the same files.
  4. Compare. Did the change fix the issue without introducing new ones?

This is especially important when you change instructions, swap context files, or adjust the output format.

Collect feedback from runners​

If your workflow is shared with a team:

  • Ask runners which findings they ignore or override — those may be noise you can eliminate.
  • Ask which findings they wish the workflow caught — those are gaps in your instructions or context.
  • Check whether users are adding the right files to the right slots.
  • Look at usage analytics to see if the cost per run matches the value delivered.

Know when to split a workflow​

If a workflow is getting long, slow, or inconsistent, it may be doing too much. Signs you should split:

  • Instructions are longer than a page.
  • The workflow checks multiple unrelated things in one pass.
  • Outputs mix different audiences (e.g., client-facing findings with internal notes).
  • Run cost is high relative to the value of any single finding.

Create focused workflows for each distinct task and let users run the ones they need.