Jesus Rodriguez
AI to Widget is an open-source toolkit designed to turn any existing web application into one with its own conversational agent — grounded in the app's real data, able to take real actions on behalf of the user with thei
Core trick
A postmortem of a failed build, categorizing 20 incidents by failure type with commit hashes and verbatim evidence.
Description
AI to Widget is an open-source toolkit designed to turn any existing
web application into one with its own conversational agent — grounded
in the app's real data, able to take real actions on behalf of the
user with their own permissions, deployable with a few Claude Code
slash commands.
The product was not shipped. Instead, this submission is a structured
postmortem of the 5-day build attempt using Opus 4.7 and Claude Code
in a spec-driven, constitution-first agentic workflow.
The deliverable is a documented report of 20 incidents across 8 failure
categories — with timestamps, commit hashes, verbatim evidence, and
honest attribution (model error, user error, ambiguous spec, or mixed).
It includes 7 actionable recommendations for Anthropic at the model,
tooling, and workflow guidance layers.
The hackathon asked builders to push Opus 4.7's limits and find them.
This report does exactly that — not as a complaint, but as data.
README excerpt
# AI-to-Widget — A Hackathon Postmortem with Claude Opus 4.7 > **What this repository now is:** the structured, evidenced report of an > entry to the *Built with Opus 4.7: a Claude Code hackathon* competition that > did not ship. The original engineering work — nine Spec-Kit features, a > constitution, dozens of artefacts — is preserved as the source material the > report draws from. ## TL;DR A solo developer entered the **"Built with Opus 4.7: a Claude Code hackathon"** with **\$500 in API credits** and a project to ship. Over five days the developer attempted nine Spec-Kit features for an open-source widget toolkit (`ai-to-widget`). A first end-to-end version landed Wednesday night (mostly built on the author's personal Claude subscription, before/around the time the credits arrived). From there each iteration introduced more regressions than it fixed. On **Saturday night, 2026-04-25**, after discovering that `/atw.build` had been silently failing for ~17 hours on top of a 19-commit branch-base mistake, the developer decided not to submit. **On Sunday 2026-04-26** he decided to turn the experience into something useful: this report. Most of the credits are gone. The data is not. This report is what came out of the attempt. The full, navigable report lives at **[`index.html`](https://chuslanzarote.github.io/ai-to-widget/)**. If this repository is published via GitHub Pages, the report is also available at the repository's Pages URL. ## Why this report exists The auth