Prompting for quality

How you ask is the single highest-leverage thing you do with Vizspect. The assistant is good at turning a clear description into a clean figure, and it works against you if you try to micromanage the placement and styling it already handles. One idea runs through everything below: describe intent and structure; let the tool handle placement and house style; iterate from renders.

Describe the figure, not the geometry

Say what the parts are and how they relate, not where they sit or how big they are. Layout, spacing, alignment, margins, and collision-aware leader lines are the tool’s job, and it does them better than a list of coordinates would.

  • Effective: “A CONTROLLER sends commands to three identical VALVES; each valve feeds a separate TANK.”
  • Less effective: “Put a box at 2,3 sized 1.5×1, another at 4,3…”

Describe relationships (A feeds B; this is a decision with a yes and a no path; these three are variants of one part) and let Vizspect work out the arrangement.

Let house style be automatic

You don’t choose the styling in the prompt; Vizspect applies the configured house style to every shape: a single font (Arial by default), all-caps text, black-and-white line art with no fill, 1.0 pt strokes, and USPTO margins. Asking the assistant for a colour or a fill won’t work at all; figures are black-and-white line art by design. Asking for a different font or text case won’t work through the prompt either, but those you can change: set them once in Style settings (tray) and they apply to everything; see House style & USPTO conventions. Let your prompts focus on the figure, not its formatting.

Lean on the source material

The assistant can read far more than a one-line description, so give it the source, and the figure stays faithful to the invention: the parts, their relationships, the element names, and the reference numbers come from your material, not from guesswork.

Three ways to do it, simplest first:

  • Paste it in. Drop in the relevant part of the invention disclosure, the draft specification, the claim language, or your notes, and ask it to draw from that.
  • Point it at a folder. If your assistant can read files (ChatGPT Codex, Claude Code, and Cowork all can), skip the pasting: aim it at a folder of context (the disclosure transcript, the disclosure documents, the draft application, earlier figures) and it reads what it needs. The richest way to work.
  • Name a claim. Ask for “a flowchart for claim 1” and it turns a method claim’s steps into a START → step → END flow you then refine. (Method claims map onto flowcharts; apparatus and system claims onto block diagrams.)

This pays off in a few ways:

  • Names and numbers match the spec. If your specification calls it a “THERMAL REGULATOR (210),” the figure uses the same label and number, drawing and text agree.
  • Structure follows the description. A paragraph on how the components interact becomes boxes and connectors.
  • Consistency across a set. Give it the earlier figures (or their element lists) when you draw a related one, so shared parts keep their numbers.

No document to hand? Even a sentence about what the system does (“this is the water path of a coffee maker”) beats “draw five connected boxes.”

Reference numbers and figure numbers

Tell the assistant which figure you’re drawing (“this is FIG. 3”) and it numbers accordingly (300, 302, …). You can supply specific numbers or let it allocate them; it follows the house scheme, including letter suffixes for repeated parts (three sensors become 320A, 320B, 320C, not 320/322/324). If you’re matching an existing figure, give it the numbers to reuse so they stay consistent.

Iterate from renders

The most reliable workflow is a short loop: ask → have the assistant render the page → look → give one specific correction → repeat. Concrete, one-thing-at-a-time feedback (“move 104 left of the pump”; “route the return line along the right edge”) beats a long up-front specification every time. Seeing the page also lets the assistant run its own geometry check before you do, catching overlaps and out-of-margin placements early.

Lean on the high-level builders

For common figure types, ask for the whole thing in one go: “build a flowchart with these steps…”, “a swimlane chart with lanes A, B, and C…”. These bundle dozens of conventions (sizing, spacing, START/END, decision branches, numbering) into a single step, so you get a coherent figure instead of a pile of boxes. When a figure looks finished, ask the assistant to run a compliance check on the page, a quick pre-render check for overlaps, stray lines, and margin problems.

Co-edit freely

You don’t have to do everything through the assistant. Drag a box in Visio, fix a label, then tell the assistant to carry on; it re-reads the current state before every change. The one rule: take turns. Visio is single-threaded, so don’t edit at the same moment the assistant is working; let each side finish.

Tell the assistant when you’ve changed something

The assistant re-reads the shape it’s about to touch before each edit, but it doesn’t watch Visio over your shoulder. It can’t see a change you made a few edits ago on the other side of the page, or on a page it isn’t working on. So whenever you’ve done hand-editing (moved boxes, renamed a component, deleted or added a figure, drawn something yourself), say so before you hand back:

“I’ve rearranged FIG. 2 and renamed the controller. Take a look before you carry on.”

That one sentence prompts the assistant to re-render the page and re-read the live document, refreshing its picture of what’s actually there. It then draws against reality instead of a stale snapshot from earlier in the conversation: reference numbers land in the right places, it reuses the labels and numbers you actually used, and it won’t collide with or overwrite the work you just did. The bigger the change, the more it pays off. After a substantial manual reorganization, a quick “I’ve reworked this page, please re-read it” is the difference between the assistant guessing and the assistant knowing.

You don’t have to narrate every nudge; the pre-edit re-read catches small local tweaks. It’s the larger, off-to-the-side edits (a new page, a moved element, a renamed part, a figure you deleted) that are worth the heads-up, and they’re exactly the ones that quietly derail a drawing when the assistant doesn’t know about them.

Worked examples

A block diagram

  • Vague: “Draw a diagram of a thermostat.” The assistant has to guess the parts and the layout, and you’ll spend several turns correcting it.
  • Effective: “FIG. 1: block diagram of a smart thermostat. A TEMPERATURE SENSOR and a USER INTERFACE both feed a CONTROLLER; the controller drives a RELAY that switches an HVAC UNIT. Lay it out left to right with arrows between stages, and add reference numbers.” You get a clean first draft you only fine-tune.

A flowchart

  • Vague: “Make a flowchart for the startup process.”
  • Effective: “FIG. 4: flowchart of the device’s startup. START → SELF-TEST → a ‘TEST PASSED?’ decision. If yes, ENTER READY STATE; if no, FLASH ERROR LED and loop back to SELF-TEST. Number from 400.” The builder lays out the steps, the decision diamond, the yes/no branches, and the loop-back for you.

When to start a new figure

Keep one figure to one idea. If you’re adding a second, unrelated diagram, ask for it on a new page (“start FIG. 2 on a new page”) rather than crowding the first. For a set, describe them together (“I need three figures: an overview, a controller detail, and the startup flowchart”) and let the assistant lay out the pages.