Bootstrap AI tooling into any repo. Then get out of the way.
One command drops a set of bootstrap skills into your project. They interview you, read the codebase, and write an AGENTS.md, project skills and sub-agents tuned to your stack, wired for Claude Code, Codex, Cursor and Copilot at once. Then they uninstall themselves.
How it works
Install. Bootstrap. Uninstall.
The bootstrap skills are one-shot. They scaffold the project once and come back out, so they never clutter the slash menu and can never overwrite hand-tuned output on a later run.
Install the skills
A dependency-free CLI symlinks the vinta- skills into whichever vendor directories you pick. Node 18 is the only requirement.
$ npx vinta-ai-workflows install --tool all
Run the orchestrator
It reads the repo, interviews you for what the code can't say, then writes AGENTS.md, skills, sub-agents and the per-vendor wiring. Read-before-write, never a blind overwrite.
> /vinta-bootstrap-ai-tools
Commit and remove
Review the generated layout, commit it, and remove the builders. Later upgrades arrive through sync, not another bootstrap.
$ npx vinta-ai-workflows uninstall --tool all
What lands in your repo
One canonical layout, every vendor sees it.
ai-tools/ is the single source of truth. A generated setup script projects it into each tool's native location, so Claude Code, Codex, Cursor and Copilot read the same skills, the same agents and the same instructions.
- AGENTS.md at the root, with
CLAUDE.mdandcopilot-instructions.mdsymlinked to it. - Skills derived from your stack, plus a project-agnostic foundation set.
- Three foundation sub-agents: implementer, reviewer, fixer.
- A config file recording every opt-in so sync stays faithful to your choices.
The workflow after bootstrap
From a raw prompt to merged code, in three skills.
Spec, plan, execute. Each step writes a durable artifact the next one reads, and every one of them interviews before it drafts.
create-spec
Never turns a vague prompt into a plausible-sounding spec by guessing. It interviews for business context, hypothesis, use-cases, acceptance scenarios, negative scope and risks, then writes the contract.
plan-feature
Translates the spec into phases with explicit dependencies, a touch list, a crew of implementers with tiers, and the feature-flag decision. Writes a markdown plan for people and a workflow.json for machines.
implement-plan
A conductor that runs independent phases at the same time in worktree lanes, drives gates to green, loops a reviewer one tier above the implementer until it approves, and opens stacked PRs.
Foundation skills
Project-agnostic skills, shipped verbatim.
These land in every bootstrapped project. Stack-specific skills come from your team's own templates, pointed at during the interview, never from a bundled one-size-fits-all body.
handoff always
A session-continuation doc a fresh session or a teammate can resume from, gathered from the repo and never from memory.
thermo-nuclear-review-loop always
One reviewer sub-agent, verified findings, fixes by the implementer, repeat until explicit approval.
open-pr-from-context always
A durable PR-context file published as a real PR with inline comments through gh or glab.
deslop-comments always
Rewrites touched comments into Simple English and deletes the ones that should not exist.
write-unit-test default
Six framework-agnostic rules plus your captured conventions, so tests assert something and leave nothing behind.
prepare-worktree opt-in
A fully runnable worktree per parallel phase, its own deps and database, with an OS-level write guard.
systematic-debugging opt-in
Root-cause-first debugging with your repro commands and observability MCP servers.
add-one-off-script opt-in
Backfills and cleanups on a base class that enforces dry-run, idempotency and segmented backups.
Also: create-qa-use-cases, add-e2e-test, add-env-var, thermo-nuclear-code-quality-review, handoff-to-client, interview-ui.
Staying in sync
Your project keeps getting better without re-bootstrapping.
New foundation skills, sharper templates and schema additions land upstream all the time. One command from your AI tool reads the changelog since your recorded version, classifies each change against what you enabled, and gates every write on your answer.
- Per-change Apply / Skip / Show diff. Nothing is clobbered.
- Opt-outs are sticky across runs, recorded in
.vinta-ai-workflows.yaml. - Schema migrations are automatic, with the diff shown first.
- Hand-authored skills are detected and left alone.
Two seconds to install. One afternoon to bootstrap.
Pin a version for determinism, or install from a git ref to try an unreleased build. Node 18 or newer is the only prerequisite.