← All projects

Workflow Assessment Tool

Describe a workflow and see it broken down through my six-part decision framework.

Most AI projects struggle before model selection because the workflow and decision are still vague. This tool turns a plain-language process description into a structured assessment of tasks, bottlenecks, prediction opportunities, human checkpoints, risks, and success measures.

Designed for controlled use

User input
→
Input validation
→
Rate-limit check
→
Restricted Anthropic request
→
Structured-output validation
→
Workflow assessment
  • ·Restricted to workflow-assessment requests
  • ·Rejects unrelated requests
  • ·Maximum input length (4,000 characters)
  • ·Structured response format, schema-validated before it ever reaches the page
  • ·Calls limited by IP address
  • ·API credentials remain server-side
  • ·Clearly labeled live or simulated output
  • ·No persistent conversation memory

What you'll get

  • ·The tasks it found in your workflow
  • ·Likely bottlenecks
  • ·Where prediction could help
  • ·A verdict of automate, redesign, or keep human-led
  • ·Risks to watch
  • ·Where a human should stay in the loop
  • ·Suggested success metrics

Not sure what to type? Click “Try an example” on the left.

What this shows

A plain-language workflow description can contain enough information for an LLM to identify its tasks, bottlenecks, and judgment calls before a person maps each one manually. The result gives the user a structured starting point for diagnosis, and it can then be checked against how the business actually operates.

How it works

You describe a workflow in plain language. The description passes through input validation and a scope check, then goes to Claude Opus 4.8 with a system prompt built from the same six-part framework on the Approach page. The response must follow a defined structure, and the page validates every field before displaying the assessment.

Design decisions and guardrails

I gave the model one narrow job, which is turning a workflow description into a structured assessment. It cannot hold an open-ended conversation, remember previous requests, or answer questions outside that scope. Restricting the model in this way also makes the input controls, rate limits, and response validation easier to enforce.

Limitations

The assessment depends on the detail provided by the user, so it should begin an operational discussion rather than end one. The tool also cannot verify what happens inside the business because it can only reason from the workflow description it receives.

Production next steps

A real engagement would compare the assessment against workflow data, employee interviews, exception logs, and measurable outcomes. The prototype starts the diagnosis, but it does not replace operational discovery.

Technology used

The tool runs through a Next.js API route deployed as a Vercel serverless function, then calls Claude Opus 4.8 through the Anthropic API with structured outputs. Zod validates both the request and the model's response, while per-IP rate limiting controls repeated use.