2buildDocumentationGitHub
2build / DocumentationRead Markdown (.md) View source

Skills & workflows

You type a handful of commands, one per step of the 2found loop. Autopilot and Team run the other skills for you as steps of their work.

Commands you type

StepCommandStops at
Start a project from an idea/bbs:autopilot "<idea>" in an empty folder (or no idea: it asks, then offers three)A runnable repository: requirement, a prototype you approve, services scaffolded from starters, local setup, release documents, the 2build harness and docs/BACKLOG.md of paste-ready /bbs:autopilot lines. One skeleton commit; nothing is provisioned or deployed
Connect a project/bbs:setup-project, then /bbs:harness-audit2build config, QA target and project docs; AGENTS.md checked against the code
Prototype/bbs:design-ui "<screens>"A clickable prototype for you to approve
Build one ticket/bbs:autopilot "<requirement>" — feature or bug fix (auto-selects build/implement/child/verify)QA-verified change, left uncommitted
Build a whole project/bbs:team "<project>" — an autonomous product team on Orca: you confirm the spec and approve the basic design, then review the handoverA reviewed product, one commit per phase with cumulative QA, released on your go at handover; in-scope feedback fixed by the team
Check performance/bbs:perf-audit "<feature>", or no target for the whole systemA report in two parts: the owner part first (what people feel in seconds, why, numbered recommendations with cost, risk and the command to paste), technical detail below a divider (guide)
Release/bbs:release (--local, --dry-run, --prepare, --resume <tag>)Verified distribution, deploy and hypercare
Market test and launch/bbs:marketing-video "<goal>"Per-platform video files (TikTok, YouTube, Reels, X, LinkedIn) from a script a simulated panel could retell; you read the script, then post them yourself

Autopilot has two built-in workflows: builder for one ticket and new, which autopilot picks in a folder with no project (or /bbs:autopilot new "<idea>"; there is no separate command). A project may add its own under .claude/workflows/. Every stop a person reads follows owner language: what happened, where to look, the command to paste.

humanizer-vi edits Vietnamese prose (scripts, captions, docs, EN → VI translations) for AI and translationese tells; it is the Vietnamese edition of blader/humanizer and needs no bbs. The agent picks it up when it writes or reviews Vietnamese, or type /bbs:humanizer-vi.

marketing-video and agent-first-docs belong to the marketing step of the loop, not to product engineering. When 2market ships they move to its namespace and keep a bbs: alias.

Roles → commands

The start guide names five kinds of work. They are roles you play, not commands; each maps to a command you already have.

RoleWorkCommand
PrototyperTry a risky idea before building it/bbs:design-ui; a throwaway code spike is a direct agent task
BuilderA feature or bug fix through review and QA/bbs:autopilot (/bbs:team for a phased project)
SweeperSimplify without changing behaviourClaude Code /simplify, or /bbs:review-pr
GrowerCopy, conversion, a growth experiment/bbs:marketing-video, /bbs:agent-first-docs; a copy or flow change goes through /bbs:autopilot
MaintainerFailures, security, dependencies/bbs:autopilot (it investigates first); /bbs:investigate for root cause only; Claude Code /security-review

Steps the workflows run

Autopilot and Team call these on their own. Type one only for a focused job outside a workflow.

NeedSkill
Plan without coding/bbs:plan-draft
Implement a scoped change/bbs:implement
Add missing behavioral coverage/bbs:test create
Audit weak, redundant or unrun tests/bbs:test audit
Measure and optimize test lanes/bbs:test optimize
Root-cause a bug/bbs:investigate
Baseline and verify a performance change/bbs:perf-audit (verify <report>)
Focused browser check/bbs:browse
Full test/fix browser loop/bbs:qa
Pre-landing code review/bbs:review-pr

semantic-decision makes bounded judgments inside the workflows (the current LLM by default, optional Cloudflare inference). It sets user-invocable: false, so it is not in the / menu.

Occasional commands

These set disable-model-invocation: true: they are not in the model's skill listing, so the agent never picks them up on its own — type the command.

NeedSkill
Create a pull request after human review/bbs:create-pr
Classify and unblock a stalled/BLOCKED run/bbs:triage
Turn 2build's own telemetry into ticket-ready findings/bbs:telemetry-review
Find where session tokens and background context go, rank savings/bbs:token-audit
Audit a running app's UI against the UI bar: broken or duplicated flows, complexity, text, component and loading drift, UI speed; inventory work flows/bbs:ui-audit
Map the current architecture and design the next generation: diagrams, headroom, a scaling roadmap with measured triggers; new designs the smallest system for a product with no code yet (guide)/bbs:architect
Evaluate external code/bbs:recon
Docs, quick start or marketing page built around one agent prompt; agent-readable Markdown/bbs:agent-first-docs
Draw an image or cut one out locally (Qwen Image, RMBG-1.4; no account); marketing-video's mvf art uses it/bbs:image-gen

Renamed, old commands still work: release-prd → release, analytics-review → telemetry-review, foreman → team.

Skill boundary

release reads the target repo's release profile (RELEASE.md, .babysit/release.md, or the linked inline instructions). Use /bbs:release --local when CI quota is exhausted: checks and builds run locally, required artifacts are published manually, and affected automatic workflows are guarded/restored. /bbs:release --local --dry-run reports the plan without mutations; --prepare stops before remote delivery and --resume <tag> recovers a partial release. Options are skill instructions, not a new CLI subcommand. Codex uses $bbs:release with the same options. Autopilot's existing QA checkpoint does not grant release authority.

A standalone skill earns its place by owning a 2build contract, evidence format, repo convention or workflow boundary. Ordinary reasoning, brainstorming, copy and script writing are direct agent tasks.

Removed shortcuts: conversion-fix, copy-rewrite, growth-experiment, office-hours, reason, social-content. Ask the agent directly for those jobs, and use agent-first-docs for the product activation/docs standard.

Removed with the archetype model: the prototype, sweep and maintain skills and the prototyper, sweeper, grower and maintainer workflows (telemetry: builder ran 396 times, the four others 7 in total). A spike is an ordinary agent task; simplify with Claude Code's /simplify or review-pr; audit security with /security-review; fix a bug with autopilot (investigate inside it). fix-pr is gone too — ask the agent to address the PR's review comments; product-marketing-page merged into agent-first-docs. The reasoning benchmark remains archived under tests/reason-bench; it is not an installed skill.

create-pr retains PR and ticket-state protocols; telemetry-review retains telemetry interpretation; triage retains bounded checkpoint recovery. Low personal usage alone is not a reason to remove these contracts.