Independent review with no Builder context. Hard gates where the run stops. Stop for consequence, never for ambiguity.
Capstan is a Claude Code plugin that takes work from concept to delivery. Five roles carry it: a Scout researches, the Architect, which is your own session, interviews and plans, each Builder builds one slice, a Reviewer checks two axes independently, and a Courier packages what ships. Nothing polls or runs unprompted. BytesNation publishes it from the MokeyBytes repository as capstan@bytesnation.
The glossary defines every word here; DESIGN.md holds the reasoning.
claude plugin marketplace add MokeyBytes/capstan
claude plugin install capstan@bytesnationRestart Claude Code; agent definitions load at session start. This installs at user scope. Add --scope project for a shared repository.
For a GitHub Projects v2 board instead of a plain file, install the board plugin too:
claude plugin install capstan-board@bytesnationOptional; see tracker surface for the cost.
Run /capstan:setup to choose where the glossary, decision log, decision records and tracker file live. See document home. Skip it, and your first effort offers the default, or a stop to run setup.
/capstan:effort add rate limiting to the public API
You talk to the Architect throughout. It owns the interview, spec, slice graph and decision log, and never builds or reviews.
- It interviews you, in rounds, each carrying a recommended answer.
- Gate one locks the concept: what, why, and what you are not building.
- Phase two plans slices and stops at gate two with what it assumed.
- Phase three builds and reviews each slice, test-first, then runs your repository's checks against the merged result.
- Phase four delivers: the Courier packages the output and writes the permanent note; you commit it once review passes.
For anything smaller, /capstan:quick runs one slice through a single gate; see disciplines.
| Gate | The brief answers | You decide |
|---|---|---|
| 1. Concept locked | What we are building, why, what we are explicitly not doing | Right thing? |
| 2. Plan locked | How, cut into slices, what runs parallel, what was assumed | Right shape? |
| 3. Ready to deliver | What was built, what review and verification found, what goes to whom | Ship? |
The run ends at every gate. That is what makes the gate real. .capstan/effort/CLAIM.md records where the effort got to, so the next session resumes there. A lock beside it refuses a second session on the same effort and names who holds it; only you decide to resume or take over.
Unclear requirements never stop the run. The crew takes the most defensible reading, logs the assumption, and keeps going. Every assumption surfaces at the next gate, where correcting one costs almost nothing. Four things do stop it: secrets and credentials, anything a third party will see, anything that costs money, and anything destructive or production-facing. The one delete the crew makes on its own is the gitignored scratch, at delivery.
Yipadoo, a paid task-management platform, was built end to end with Capstan. Every feature went through the interview, the three gates, independent review and delivery. Its code is private; the product is live.
Four front doors, invoked only by you, none a discipline: effort starts a full run, quick runs one slice through a single gate, setup configures where the artifacts live, and agent-models tunes the crew's model and effort for Claude and Codex. Installing capstan-board adds a fifth, /capstan-board:setup, for the tracker surface.
| Skill | For |
|---|---|
interview |
Rounds of questions, each with a recommendation. |
spike |
A throwaway build for a stalled design question. |
slicing |
Cuts a locked plan into vertical slices with blocking edges. |
test-first |
Red, green, refactor, tested only at agreed seams. |
diagnosing-bugs |
A feedback loop that goes red on the bug before any theory. |
codebase-design |
Words for structure, so a review can call a module too shallow. |
two-axis-review |
Standards and spec, reviewed independently, never blended. |
verify |
Runs the checks your repository declares against the merged result. |
resolving-merge-conflicts |
Integrating parallel Builders, where neither Builder can be asked its intent. |
walkthrough |
The one-time script for a manual procedure. |
decision-record |
A one-line log by default, a full record only when earned. |
brief |
Checkpoint and partner briefs, per recipient, never maintained. |
to-questionnaire |
Turns a question nobody present can answer into a document for whoever can answer it. |
unslop |
Cuts AI tells from prose a person reads. |
writing-for-agents |
Keeps a document an agent consumes flat and consistent. |
A real brief goes here once bench/ produces one. Nothing here is invented; see bench/PROTOCOL.md for how one gets measured.
Models and effort come from each agent's frontmatter in agents/*.md: Builder runs opus/high, up to capstan-max-builders (3) at once. Reviewer runs opus/xhigh, once per slice and again per fix dispatch, up to capstan-max-fix-dispatches (5), plus once more on the knowledge-base note when configured. Scout and Courier both run sonnet/medium; the Architect is your own session, not a subagent. Run agent-models to retune these when a new model ships.
What scales the bill: slices cut, fix dispatches per slice, and Scouts fired. A one-slice quick run costs one Builder and one Reviewer, plus up to two fix dispatches, each reviewed by a fresh Reviewer.
No dollar or token figure appears until bench/ measures one. See bench/PROTOCOL.md for the benchmark against plain Claude Code.
The GitHub tracker board moved out of core into its own plugin, capstan-board. If capstan-tracker already names a GitHub project, install capstan-board@bytesnation before your next effort or quick. Otherwise an effort stops before its interview, and a quick after its interview: surface not installed: install capstan-board@bytesnation. A project that has never set capstan-tracker sees no change. See upgrading and tracker surface for the rest.
/capstan:effort cannot be invoked by a model. It carries disable-model-invocation: true, same as quick and setup, so only you start it.
The effort: setting is not supported on Haiku. Drop the effort: frontmatter line from any agent you point at a Haiku model.
Bash is an escape hatch. Builder, Reviewer and Courier all hold Bash, so their "never do X" rules are prose, not enforcement. examples/ ships a settings.json deny list, but it matches text, not the program, so it can't stop /bin/rm -rf. Builder's acceptEdits skips file-write prompts; Bash still prompts, stalling fan-out.
Fan-out does nothing for single-artifact work. Parallel Builders need slices owning different files; a document, a video script, or a config file is one artifact and one Builder.
The helpers refuse; they do not intercept. capstan-claim refuses a second owner, a sixth fix dispatch, or a fourth Builder. Three efforts at once is the ceiling. The Architect enforces it by asking which to close first; no helper counts efforts. Nothing stops an agent from running rm -rf or spawning a Builder by hand. The tests prove what each helper does when called, not that every agent calls it.
Reading a live board and tearing it down has been tested once, for real. A live trial moved 42 items off a GitHub board: every item read back, the reconstruction diffed clean, every item removed and the open milestones closed. Moving onto a board, in its current form, has not run end to end against a real one. Its writes ran live under earlier releases, and the read-back check that now guards the delete has run live only on the way off.
The knowledge-base note is reviewed whole, when there is one. Capstan never commits it, so it has no fixed point to diff against.
More: document home and vault layouts, tracker surface, knowledge base, upgrading and moving the marketplace, installing by hand, what CI runs.
MIT. See LICENSE. Take it, change it, ship it.
Some skills here are not ours, most from Matt Pocock, MIT and redistributed with a CREDIT.md recording every change:
skills/writing-for-agents/:SKILL.md,SKILL-MECHANICS.md.skills/unslop/:SKILL.md, by Lauren Tan via cursor/plugins.skills/walkthrough/:template.sh, forked.skills/diagnosing-bugs/:SKILL.md.skills/codebase-design/:SKILL.md,DEEPENING.md,DESIGN-IT-TWICE.md.skills/to-questionnaire/:SKILL.md.skills/resolving-merge-conflicts/:SKILL.md.
More ideas come from him: the frontier in interview, from grilling; the next field in CLAIM.md, from handoff; the unformed log status, from wayfinder; two interview moves, from domain-modeling.
