Status	Last updated	Parent	Read before this
complete	2026-04-30	none	none

Vision

Current state

devwork-skills exists to make high-volume, high-quality feature development the normal output of a single person working with Claude Code. The aspiration is the kind of throughput Boris Cherny reported — feature after feature shipped, with Claude writing the code — made repeatable across projects and tech stacks.

The shape of the work is a product-owner model. The human's role on a feature is three decisions: approve the requirements (PREQ), approve the technical approach (plan), and accept the result (UAT). Everything between those decisions runs without intervention.

Repeatability matters more than peak performance. A solo founder, a tech lead with a small team, or a developer running multiple projects in parallel should all be able to point Claude Code at a feature and trust it to come back with something worth reviewing. The system optimises for "consistently good" over "occasionally brilliant."

Rationale

The three-gate model is the result of asking, for each phase, whether a human decision is genuinely required or just habitual. Requirements are a value judgement — only the product owner can say what's worth building. The technical approach has consequences the system can't reason about alone (rollback risk, cross-feature coupling, organisational fit). UAT confirms the thing works as desired in the real environment. Code review, test design, security review, accessibility review — those are mechanical enough to delegate, with the right structure.

Optimising for repeatability over peak rules out heroic flows that only work when the operator is paying close attention. It also rules out optimisations that depend on a specific stack or forge — the system has to work on whatever the project happens to use. This is what drives the Contract - Forge Adapter and the stack-agnosticism in the skills themselves.

Open questions

None at this level. Specifics belong to lower pages.