Skip to content

Roadmap

This site documents only what’s merged and runnable. Everything below is planned but not yet built on the current codebase — rather than scatter half-true promises across the concept and reference pages, they’re collected here once, pointing at the actual planning document.

The full, current build plan (phase scope, ordering, and verify recipes) lives in the product repo at docs/_archive/plans/2026-07-10-playmaker-to-skillmaker-migration/plan.md. A running narrative of what’s actually shipped, phase by phase, is in the build log at docs/_archive/plans/2026-07-10-playmaker-to-skillmaker-migration/build-log.md.

Recently shipped (see the concept/CLI pages, not this list)

Section titled “Recently shipped (see the concept/CLI pages, not this list)”

Agent-first production stations, publishing, the skillbook, a second provider (codex), and brownfield adoption of existing skills repos all shipped since this page last listed them as “coming.” They’re documented where the rest of this site documents merged, runnable features, not here:

  • Agent-first production stationsskillmaker station run, backing William’s own skill-drafting. See The Skill Bundle (stations.json) and the build log.
  • Publishingskillmaker publish, git-dir and marketplace targets.
  • The two-audience publish door, in the viewer too. skillmaker publish <slug> --to user|project installs a recorded version’s output/ where an agent actually reads it (~/.claude/skills/ or the project’s .claude/skills/), remembers the audience in bundle.json, and supports revert to an older snapshot via --version <hash>. The viewer’s Publish tab drives the same core through live Publish/Revert buttons — they are no longer disabled.
  • The skillbookskillmaker book build, and Publishing and the skillbook for how the two relate.
  • Codex, a full second provider — a real eval peer to claude-code, not just configured; see Running fixtures.
  • Adopting an existing reposkillmaker adopt and the getting-started walkthrough, including real numbers from adopting two real-world skills repos.
  • The predecessor studio’s library — migrated and Raven-reviewed against the shipped data model (Phase 14).
  • William, self-hosted for real (Phase 19). The repo’s own studio is William’s actual working environment: real design.mds, fixtures by failure class, a risk map, recorded versions, and measured runs on both providers, staged through the board like any other bundle.
  • The machine-level project registry. skillmaker start now serves every project registered at ~/.skillmaker-studio/config.json and ignores the directory it’s run from; projects are added with skillmaker project add/list/remove or the viewer’s New-project dialog (browse, create, and auto-init a directory). See skillmaker start.
  • Published on npm, four platforms. npm install -g skillmaker-studio ships prebuilt binaries for macOS arm64/x64, Linux x64, and Windows x64 (Windows is new and lightly tested). See Install.
  • William ships inside the product. Station skills (william-research-a-skill, design-skill, william-draft-skill-md) are packaged with the CLI and resolve as a fallback when the workspace has no copy of its own — no more hand-carrying the William bundles into every workspace. See The Skill Bundle.
  • Version content snapshots. skillmaker version record snapshots the recorded content under the bundle (.skillmaker/versions/), and version show lists it back — version history no longer depends on git. See Versions and drift.
  • Phase 20 — six real personas, six friction logs, four fix PRs (v0.2.0 → v0.2.1). Six fresh-install stories (first bundle, model selection, vendored-adopt, publish/storefront, maintain/re-earn, failure-iteration) each drove real friction logs against the shipped CLI, closed by PRs #47/#49/#51/#54: layout-aware runs, journal safety, sandbox isolation, --model selection with resolved-id recording, storefront receipts, response.md, honest confidence intervals, run repair, adopt --source provenance, and — the story’s sharpest finding — a credential-leak P0 in sandboxed auth, fixed structurally in v0.2.1. See the build log for the full phase entry and Provider auth & troubleshooting for what shipped.
  • Desktop app: Windows/Linux, a signed downloadable build, and registry-era rework. The macOS build-from-source shell is documented in Desktop app; a cross-platform, installable version is still ahead (the CLI now ships prebuilt for four platforms via npm, but the desktop shell doesn’t), along with updating the shell’s per-workspace picker/claim logic for the machine-level project registry and an in-app reconnect for the “attached-but-dead server” limitation noted there.
  • Iteration ergonomics: a runs list command and grading at volume. Story 6’s friction log found real pain grading 8+ runs in a row: no skillmaker runs list (run ids currently live only in dir names, and must be copy-pasted from run output), and no --fixture ... --latest or similar selector to queue grading without hand-tracking which run is which. A -k/count-oriented flag for driving several runs of the same fixture in one invocation (instead of N separate skillmaker run calls) is in the same backlog bucket. Not yet built.
  • Drift vs. upstream. adopt --source/--ref (shipped in v0.2.1) records where an adopted skill came from at adopt time, but the product still can’t answer “has the source repo changed since I adopted?” — a skillmaker upstream diff (or similar) that re-fetches --source and compares would close that. Today “drift” only means drift vs. your own adopt-time snapshot, not vs. upstream.

Every other page on this site was written and verified against the real, running CLI and viewer on this branch — including pasted command output from an actual fresh-directory walkthrough. Nothing in What’s coming exists to run yet, so writing detailed pages for it would either go stale immediately or read as marketing rather than documentation. This page will be broken up into real sections as each phase ships — most recently, that’s exactly what happened to agent-first stations, publishing, the skillbook, codex, and adopt.