Skip to content

mylesmcook/product-wrangler

v0.2.1MIT

Turn an existing runnable web app into an evidence-backed, complete product without expanding authority.

GitHubOpen Website

Product Wrangler

Product Wrangler is a small, framework-neutral, skills-only plugin for turning an existing runnable web app into the product it should be.

It is intended for broad implementation requests such as:

$product-wrangler wrangle this existing app into the product it should actually be

This app works, but it feels half-baked. Figure out what is wrong and fix it end to end.

It does not own greenfield application building, isolated UI fixes, review-only critique, or backend-only work.

Contract

Product Wrangler requires the agent to:

  1. Inspect the dirty worktree, rollback boundary, runtime endpoints, data class, test accounts, and outbound effects before mutating browser flows.
  2. Bound one primary product surface instead of expanding through the entire app.
  3. Infer the product from a ranked evidence hierarchy and attempt to disprove the selected thesis.
  4. Record contradictions, assumptions, confidence, non-goals, and evidence for every major product decision.
  5. Create a workflow ledger that ties each primary user outcome to baseline behavior, an expected observable result, a selected change, post-fix evidence, and remaining uncertainty.
  6. Replay the same journeys before and after, run a cold-start pass, revisit the thesis, and select COMPLETE, PARTIAL, or BLOCKED honestly.

The plugin does not add tools, hooks, MCP servers, framework dependencies, or operational authority.

Safety model

A local frontend may still point at production. Product Wrangler identifies backend endpoints and external effects before testing. Mutating browser journeys run only against fixtures, disposable accounts, isolated local data, or proven sandbox services. Unknown or real-service-backed applications are inspected read-only unless the exact side effect is authorized.

Behavioral evaluations

evals/evals.json covers broad triggers, nearby non-triggers, conflicting product evidence, generic redesign, dirty worktrees, real-service-backed apps, missing browser or credentials, cosmetic-only iteration, scope growth, authority expansion, and premature completion.

The release gate in evals/README.md requires real model runs. Manifest and contract validation alone are not evidence that the skill makes sound product judgments.

Status

Available from the mcook-plugins-local marketplace. The initial release gate included two-family matrix sweeps plus disposable, real-service read-only, and large-application scope tests. See evals/results/2026-09-04-v0.2.0.md.