An unconstrained coding agent produces code that looks right — plausible, large, and wrong in ways you only find at review. BigIn Skills adds the structure that makes AI work reviewable: an approved spec before code, an independent verifier that never reads the implementer's own summary, and commit gates that can't be switched off.
Each one has a mechanism pointed at it — not a prompt asking nicely.
You get a large diff built on an assumption nobody stated, and you find out at review.
PLAN.md, enforced by a hook, before any non-trivial edit lands."I implemented the feature and it works" is a self-report, not evidence.
Every session re-derives the same context, and re-derives it differently each time.
The loop every non-trivial task runs through. Most of it is automatic — you answer twice.
One sentence, before any code — then which rung: task, epic, or discovery. Only the first continues here.
What changes, edge cases, security, and an explicit not in scope.
The approved spec plus a task table, on disk. The working contract.
Routed to the cheapest tier that can do the job — quick, standard, or deep.
A fresh agent audits the diff against the plan. On a fail, back to implement. Three rounds maximum.
A human pass, /code-review, or both — then distil what's durable and archive the plan verbatim, so the reasoning survives.
Seventeen skills and seven subagents. You invoke almost none of them by name — they trigger on intent. If you would rather ask than guess, /ask-bigin names the one that fits and hands off.
bigin-harness-setup scaffolds a CLAUDE.md brief, path-scoped rules, and commit-time gates into an existing repo — or, on an empty one, scaffolds the app first and layers governance on top. Six stack profiles, and it generates the Cursor mirror when teammates work there.
task-workflow takes one task from spec to verified diff; epic-workflow sits above it, splitting an initiative into shippable units and feeding them back one at a time; discovery-workflow sits above both, turning a vague ask into the approved brief and PRD they assume already exists — plus the architecture decisions the product forced, written into knowledge/. debug-workflow supplies bug triage, and write-tests writes either unit tests for one function or an end-to-end spec straight from a PRD acceptance criterion.
model-router scores capability and verification separately, then routes to the tier that fits. Three ladders, from cost-first to frontier.
knowledge-distill pins a library's docs at an exact commit and audits the result. sprint-distill compresses what the team learned — never appends. Finished plans and epics are archived verbatim as an append-only record log, so why a change took its shape outlives it.
New-project generators for Nuxt 4, Next App Router, Go (Gin + GORM), and Fastify — each wired to the same harness on the way out.
Deterministic Node guards: spec gate, Conventional Commits, a regression test on every fix, and a three-stage prompt-injection gate with a canary. One guard body serves both hosts — Cursor teammates get the same gates off the same files, generated from CLAUDE.md and never hand-written.
Every rule here is enforced by something that isn't the model. A rule an agent can talk itself out of is a suggestion — and suggestions decay.
Why the harness exists, the five concepts behind it, setting up a repo, the daily loop, and the practices that decide whether it works for a team. One page, no signup.
/plugin marketplace add tammai/bigin-skills
/plugin install bigin-skills@bigin