Skillmaker Studio

Ship agent skills with receipts.

Skillmaker Studio turns SKILL.md from a paste of unknown pedigree into a measured, versioned, human-approved product — research, design rationale, eval fixtures, and pass rates included.

View on GitHub See how it works →

install

$ curl -fsSL https://skillmaker.studio/install.sh | sh

Pre-alpha · v0.2.1 · macOS arm64 + Linux x64

Design is a first-class artifact.

Every skill carries a design doc: intent, triggers, the workflow, failure hypotheses. Six months later you'll know why the skill is shaped the way it is — and so will the next person.

Evals with honest math.

Fixture cases by failure class — golden, refusal, empty-input, rerun, hard-case — mapped against five risk families. Coverage (“a fixture exists”) and validation (“passes at 28/30 on claude-code”) never merge. A single run is a sample, not a measurement.

A production line, not a folder.

Skills move idea → researching → drafting → evaluating → published through guarded gates: forward requires an approved review, publishing requires the gate, and moving backward is legal — regression is a modeled fact. Every decision lands on an append-only, git-tracked journal.

How it works

Step 1 skillmaker new
Scaffold a Skill Bundle.

skillmaker new <slug> creates a design doc, fixture slot, output slot, and stations for the skill — a real bundle, not a bare file.
Step 2
Build and prove.

Draft SKILL.md from the design, add fixtures by failure class, run them against real agents, grade the read-outs.
Step 3 skillmaker advance
Move through the gates.

Advance the bundle through idea → researching → drafting → evaluating → published, with the journal recording every judgment. Ship the output; keep the asset.

skillmaker in 30 seconds

$ skillmaker init
skillmaker: initialized workspace at ~/projects/my-skills

$ skillmaker new bash-guardrails --name "Bash Guardrails"
skillmaker: created bundle bash-guardrails

$ skillmaker status bash-guardrails
slug:        bash-guardrails
name:        Bash Guardrails
stage:       idea
substate:    working
events:      1
last event:  bundle.created at 2026-07-11T09:47:31.618Z
fixtures:    0
coverage:    0 covered, 0 partial, 0 gap

$ skillmaker start
skillmaker: serving my-skills at http://localhost:4322

More under the hood

Two doors, one ground.

CLI and live board share the same journal — agents work the CLI, humans judge in the browser, nothing desyncs.

Version-pinned evidence.

Measurements bind to the content hash they exercised. Change the skill, and the numbers honestly reset to “not yet measured.”

Drift you can see.

Hand-finish your SKILL.md freely — the studio shows that it diverged from the design, and when, without policing it.

Journal-native todos.

The work-tracking lives in the same event stream as everything else. git log is your audit trail.

Graded read-outs.

Grade a run pass, fail, or partial — in the CLI or the board — and watch it join the read-out as n · pass-rate · confidence interval, per fixture, version, and provider. Never pooled.

Desktop app.

A double-clickable macOS shell around the same board and journal — no terminal required. Build it from source today; a signed download is on the roadmap.

The skillbook.

Generated docs for your whole skill set — design rationale, changelogs, and version-pinned pass rates inline, one page per skill.

Agent-first production.

Stations run over ACP: an agent researches, drafts, and evaluates; you approve at the gates. William's own skills ship through this loop.

Publish with receipts.

Send a skill's output to a git directory or a marketplace manifest once it clears the gate — the same guarded path from the CLI or the board.

Adopt an existing repo.

Already have a folder of SKILL.md files? Wrap them as bundles in place — no files moved, no rewrite required.

Two providers, one measurement.

claude-code and codex are both full eval peers — the same fixture, measured per provider, never pooled together.

Questions

I can just write a SKILL.md by hand.: You can — and Skillmaker won't stop you (hand-edits are legal, drift is displayed, not punished). The studio earns its keep the first time a skill misbehaves and you need to know what changed, what was tested, and what was approved.
Is this only for Claude?: Bundles target any ACP-drivable agent. claude-code and codex are both full eval peers today — one skill, measured per provider, never pooled together. Publishing to a Codex-format marketplace manifest is still best-effort (no published spec to conform to yet); the git-directory and Claude-marketplace targets are on solid ground.
Where does my data live?: In your repo. Files plus a git-tracked JSONL journal. SQLite is a throwaway index. No server, no account.