CloudSketch AI Logo
FeaturesTemplatesPricingEnterpriseAboutContactLog in
🌙☀️
Log inStart free
Home/Features/Accuracy benchmark

Accuracy benchmark · results as of 11 Nov 2026

How accurate are CloudSketch AI's diagrams?

We test every diagram type on prompts the product has never seen. On our latest sealed sets, 95–100% of diagrams had no real error in each of the 18 types, and a diagram that does not draw counts as an error. This is an internal, AI-graded test; here is exactly how it works, including where we are weakest.

18diagram types
40qualifying sealed sets
3,200diagrams in the qualifying sets
95–100%no real error, per type

Results by diagram type

Diagram typeSealed setsNo real error (each set)Fully correct (strict)
ERD (database)2100% · 100%90–96%
C4 (context, container, component)397.2% · 98.6% · 98.6%90–94%
Cloud deployment296.9% · 99.0%88–92%
Sequence296.3% · 97.4%86–91%
Flowchart295.0% · 96.3%92–94%
Mind map296.5% · 97.9%94–95%
State393.8% · 96.3% · 97.5% (95.8% overall)82.5–90%
Data flow (DFD)297.5% · 98.8%87.5%
Class (UML)396.3% · 96.3% · 97.5%88.8–90%
Network296.3% · 96.3%96.3%
Use case296.3% · 96.3%77.5–83.8%
Swimlane295.0% · 96.3%80–82.5%
User journey295.0% · 96.3%95%
Quadrant295.0% · 96.3%91.3–92.5%
Sankey flow392.5% · 96.3% · 98.8% (95.8% overall)86.3–97.5%
Data chart (pie, bar, line)296.3% · 97.5%90–92.5%
Git branching (gitgraph)295.0% · 96.3%81.3–88.8%
Gantt: software, IT and business project plans296.3% · 97.5%88.8–91.3%

Each set is 40 prompts, each drawn twice: once for a technical reader and once in plain language for managers (80 diagrams). Beta and not included: Gantt charts for crop seasons, festivals and events (mostly 87–92% on mixed sets).

How the test works

  1. Fresh, sealed prompts. Each set is written in advance and never used while we change prompts or code. Every new set has 10 prompts set in India, 20 from the rest of the world and 10 with no country, because local facts (payment systems, regulators, units) are where diagrams most often go wrong.
  2. The live settings. Diagrams are made with the same model, prompts and repair steps as the live site.
  3. Graded against a rubric. A separate AI reviewer grades each diagram as pass, weak or fail against a written rubric for its type. Facts that code can measure are measured, not guessed.
  4. It has to draw. Every diagram is drawn with the same Mermaid version the site uses. A diagram that does not draw counts as a real error.
  5. Two fresh sets, same code. A type qualifies only when two fresh sets on the same code both reach 95% with no real error. When a third set was run on that code, all of them are shown and must average 95%.

Where we are weakest

  • State, Sankey, Swimlane, Quadrant, Flowchart and Git branching pass with the narrowest margins (one set at 92.5–95%).
  • Strict scores are lower for Use case (77.5–83.8%) and Swimlane (80–82.5%): the diagram is right, but the reviewer often finds something that could be clearer.
  • Prompts from outside India are usually the hardest part of a set, which is why sets now carry twice as many of them.
  • Seasonal plans (farming, festivals, events) depend on real-world timing knowledge and stay in beta.
Try it on your own system →Browse 600+ templates

Questions

Who graded the diagrams?

A separate AI reviewer, given the prompt, the diagram and a written rubric for that diagram type. Where code can measure something exactly (dates in a Gantt plan, arrow directions, whether the diagram draws at all), the reviewer is given the measured facts. It is an internal test, not an independent audit.

What counts as a "real error"?

A diagram that is wrong or broken: a missing or invented part that changes the meaning, a wrong relationship, a wrong fact, or a diagram that does not draw. Small issues that do not change the meaning (a label that could be clearer) are graded "weak" and are not counted as real errors. The strict score, which counts only fully correct diagrams, is shown next to it.

Could you have tuned the product to these prompts?

The sets are sealed: written in advance and never used while changing prompts or code. A type only qualifies on two fresh sets run on the same code; a set scored before a later fix does not count.

Is every diagram type included?

All 18 types are in the table. Gantt charts for crop seasons, festivals and event plans are still in beta and are not part of the claim.

We re-run the benchmark when we change how a diagram type is made. Questions about the method: support@cloudsketchai.com.