Deployed agent

Howard

Turns raw onboarding-walkthrough footage into sourced, ruled-out UX findings — built for speed without cutting corners on rigor.

Brief

Howard evaluates B2B SaaS onboarding flows using structured heuristic analysis — Nielsen's heuristics, Hick's Law, Fitts's Law, and the Peak-End Rule — and produces a dossier per tool: a diagnosed target, a documented pass ruling out alternative causes, and a rating for likely effect on customer acquisition. Every finding is checked against independent outside signal where it exists, and every claim is written to be checkable, not just trusted.

Tools evaluated to date

Structured findings

Developed into full case studies

Architecture

Howard is not a separate autonomous system. It's a persona and instruction set — a system prompt — that Claude Code runs under. When Claude Code is loaded with Howard's instructions, it operates as Howard for that session: same pipeline, same evaluation method, same reporting format, every time.

The pipeline itself is human-relayed, not fully automated. Footage and files are delivered by hand through a terminal session — Claude Code generates the commands, and I run them. This isn't a background process checking for new work on a schedule; it's an on-demand system I operate directly.

Output

A structured dossier per run: diagnosed target, ruled-out reasoning, held-up evidence, outside-signal cross-check where available, and a rating (Greatly Help / Slightly Help / Slightly Hurt / Greatly Hurt) for each tool evaluated.

User guide

What it does: heuristic evaluation of onboarding flows, with a documented ruling-out pass on every finding — not just an opinion.
How it runs: on-demand, via a terminal session with Claude Code operating under Howard's instructions. Not an unattended daily loop.
Input: screen-recorded walkthroughs of a product's onboarding flow, delivered by hand.
Output: a dossier — findings, reasoning, rating — per tool evaluated.

Testing summary

One documented failure, not a 100%-pass summary — a real one reads as engineering, not a demo.

What broke

Howard's original storage layer routed through OneDrive via the FlowHunt MCP server. Every attempt returned a hard 400 authentication error.

Root cause

Confirmed against Microsoft's own documentation: the Microsoft 365 connector requires a work or school account tied to a Microsoft Entra tenant. Personal accounts are explicitly unsupported — not a bug, not a config issue, a structural account-type mismatch.

Fix

Removed the OneDrive dependency entirely. Footage now moves through direct terminal relay, and the dossier is produced the same way regardless of storage layer. The failure was isolated to storage — Howard's evaluation and reasoning pipeline was never affected.