Deep Skills

Coding agents fail in two expensive ways: they build the wrong thing because a requirement was misread, and they break unrelated code because nobody mapped the blast radius. deep-skills puts a process around the agent that catches both.

What's in the pack

• deep-feature conducts the whole flow from a ticket, skips what doesn't apply, and stops at both gates.
• deep-plan reads the material, maps the code, sweeps edge cases, plans tests and asks every open question with a recommended answer. It implements nothing until you say go.
• deep-test finds the edge cases that can really happen, audits the unit tests, and checks that tests fail when the behaviour breaks.
• deep-ui-test turns requirements into a scenario matrix and runs it in a real browser across states, viewports and locales.
• deep-ux walks the changed screens as a first-time user and returns ranked suggestions. It changes nothing until you pick.
• deep-review reviews the branch against the plan twice: its own pass and an independent reviewer.
• deep-ship commits only what belongs to the work, pushes safely and opens a reviewer-facing PR.
• deep-pr-comments checks each review comment against the code, fixes the valid ones and drafts replies.

How it works

Every step reads and writes one work doc, so a new or compacted session continues where the last one stopped. Claims about current behaviour point at a file and line, or are labelled as assumptions. Your repo's own rules win over the skills.

Where it stops

It never implements or ships without your explicit go. It's a heavy process on purpose: skip it for a one-line fix or a quick experiment.