Cry less. Ship more.
Turn AI-written code into shippable products.
Builders Don't Cry (BDC) is a proof-driven agent harness for indie AI builders who want real products, not fragile demos.
You bring the intent. Agents do the work. BDC keeps the work bounded, evidenced, resumable, and ready to ship.
中文 · Release · Install · Downloads · Workflow · Skills · Commands · Host Support
- AI writes fast, but it drifts from the goal.
- If you do not read code, you cannot tell whether the AI actually did it right.
- AI often says "done" without tests or evidence.
- Session context breaks when you come back tomorrow.
- PRs, CI, reviews, and async state go unattended.
- Small bugs turn into large rewrites.
- Halfway through a plan, AI starts reinventing the architecture.
- When it is time to ship, you are not sure whether you can trust the result.
BDC fixes the part after "AI can code."
SHAPE PLAN BUILD PROVE WATCH SHIP RESUME
┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐ ┌──────┐
│ Idea │ ───▶ │ Plan │ ───▶ │ Exec │ ───▶ │Verify│ ───▶ │ PR │ ───▶ │ Ship │ ─▶│Pickup│
│ │ │ │ │Loop │ │Gate │ │Watch │ │Gate │ │ │
└──────┘ └──────┘ └──────┘ └──────┘ └──────┘ └──────┘ └──────┘
idea plan execute verify watch-pr ship handoff
autoloop pickup
| Stage | Builder does | Agent does | BDC enforces |
|---|---|---|---|
| idea | explain the intent | shape it into a clear direction | no premature coding |
| plan | approve the plan | create atomic steps | plan lint, oracles, stop conditions |
| execute/autoloop | stay out of the loop | implement, test, fix, continue | step packets, evidence, scope checks |
| verify | review the evidence | inspect completed work | read-only review, Verified HEAD |
| watch-pr | wait for async state | monitor PR, CI, review, post-merge | event dashboard, action tags |
| ship | make the final call | summarize readiness | GO/NO-GO, rollback, privacy gates |
| handoff/pickup | resume later | load selected context | no auto-loading private state |
Latest stable release: v0.1.2.
Prerequisites: git, bash, python3, and jq. GitHub features also need gh.
git clone --branch v0.1.2 --depth 1 https://github.com/0xmariowu/builders-dont-cry ~/.bdc
~/.bdc/install.sh --target auto --scope project --dry-run
~/.bdc/install.sh --target auto --scope projectThen run:
./.bdc/bin/bdc doctorDefault install is project-scoped. It creates .bdc/ runtime folders, host projection blocks, and project-local skills/commands/agents, but it does not create a global bdc command unless you ask for it.
For Claude Code, project install also writes BDC-managed hooks into .claude/settings.json. Use --no-claude-hooks if you only want the CLI and projection files.
You can install from source with git clone, or download a host package from the release:
| Package | Use when | Download |
|---|---|---|
| Claude Code | You want the Claude Code plugin, commands, skills, agents, hooks, and CLAUDE.md |
builders-dont-cry-v0.1.2-claude-code.tar.gz |
| Codex | You want Codex skills, projected agents, AGENTS.md, and CLI gates |
builders-dont-cry-v0.1.2-codex.tar.gz |
| Cursor | You want Cursor rules backed by the shared CLI gates | builders-dont-cry-v0.1.2-cursor.tar.gz |
| Project runtime | You want the project-local runtime package | builders-dont-cry-v0.1.2-project.tar.gz |
The lifecycle above is the entry path. Under the hood, each stage activates plain Markdown skills plus executable CLI gates.
| Stage | Main command | Skills | Runtime gates |
|---|---|---|---|
| Shape | idea |
idea, using-bdc | bdc route intent, bdc route skill-hint |
| Plan | plan |
plan | bdc plan lint, bdc plan review, bdc plan resolve, bdc plan first-principles, bdc plan source-check, bdc plan interface-check |
| Build | execute |
execute, verification-before-completion | bdc execute, bdc step-packet write, bdc project-gate, bdc scope-check, bdc completion-claim |
| Build loop | autoloop |
autoloop | bdc autoloop --plan <plan> |
| Prove | verify |
verify | bdc verify --deep, bdc verify-personas, bdc verify-tags |
| Watch | watch-pr |
watch-pr | bdc watch-pr daemon, bdc pr-events-flush, bdc pr-watch-ensure-daemon, bdc pr-watch-backfill |
| Ship | ship |
ship | bdc ship, bdc approval-marker, bdc release-check, bdc build-dist |
| Resume | handoff/pickup |
handoff, pickup | bdc handoff write, bdc pickup --list --human, bdc startup-resume, bdc handoff-diff |
| Health | health |
health | bdc doctor, bdc hook-health, bdc projection-check |
| What you are doing | Command | Principle |
|---|---|---|
| Decide whether BDC should run | bdc route intent "<prompt>" |
Route by intent, not keywords |
| Check BDC health | bdc doctor |
No-spend local checks first |
| Validate a plan | bdc plan lint <plan> |
Every step needs proof |
| Execute one approved step | bdc execute --plan <plan> --step <F001.S1> |
One step, one evidence packet |
| Run bounded autonomy | bdc autoloop --plan <plan> |
Autonomy with failure budget |
| Review completed work | bdc verify --deep |
Independent read-only review |
| Decide release readiness | bdc ship --root <path> |
GO/NO-GO, not vibes |
| Save session state | bdc handoff write --topic <topic> |
Tomorrow should not start from zero |
| Resume selected context | bdc pickup --list --human |
Explicit selection only |
| Watch async PR state | bdc watch-pr status |
CI and review should not be unattended |
| Install Claude hooks | bdc install-claude-hooks --settings .claude/settings.json |
Automatic gates beat manual memory |
| Install PR watch service | core/scripts/install-launchd.sh --dry-run |
Background watch requires explicit confirmation |
| Build host packages | bdc build-dist |
One core, multiple hosts |
-
Idea to shipping path Shape the idea, approve an atomic plan, let agents execute, review evidence, and decide whether to ship.
-
Autonomy without drift Agents can implement, test, fix, and continue. They cannot silently change scope, architecture, or verification gates.
-
Trust through evidence BDC does not accept "looks good." Every meaningful step needs commands, exit codes, evidence, risks, and next actions.
-
Protection from false finishes The worst failure is not an error. It is confident unfinished work. BDC uses plan lint, step packets, scope checks, verify gates, and ship checks to catch it.
-
Continuity across sessions Handoff, pickup, and startup resume make tomorrow's agent know what changed, what passed, what failed, and where to continue.
-
A real product workflow PR watch, post-merge checks, auto-remediation gates, and ship checks cover the path from local work to release decisions.
Every BDC skill is a workflow contract, not a pile of tips.
SKILL.md
├── frontmatter name + trigger description
├── Contract what the agent is allowed to do
├── Required gates executable checks before progress
├── Evidence what must be recorded
├── Stop conditions when the agent must stop instead of guessing
└── Runtime commands CLI gates that prove behavior
The important design choice: skills tell agents when to stop. BDC is built for the failure mode where AI sounds confident while skipping the work that would prove it is right.
BDC has one core contract and host-specific projections.
| Host | Support | What gets installed or generated |
|---|---|---|
| Claude Code | full adapter target | .claude-plugin, commands, skills, native agents, hooks, CLAUDE.md |
| Codex | full adapter target | .codex-plugin, projected skills, projected agents, AGENTS.md, CLI gates |
| Cursor | rules-first target | .cursor/rules/bdc.mdc plus shared CLI gates |
See Compatibility for the detailed support matrix.
builders-dont-cry/
├── README.md
├── install.sh / update.sh / uninstall.sh
├── bin/
│ └── bdc # shared CLI entrypoint
├── core/
│ ├── kernel.md # cross-host behavior contract
│ ├── skills/ # BDC lifecycle skills
│ │ ├── idea/
│ │ ├── plan/
│ │ ├── execute/
│ │ ├── autoloop/
│ │ ├── verify/
│ │ ├── watch-pr/
│ │ ├── ship/
│ │ ├── handoff/
│ │ ├── pickup/
│ │ ├── health/
│ │ ├── using-bdc/
│ │ └── verification-before-completion/
│ ├── agents/ # read-only verification personas
│ ├── standards/ # plan, evidence, routing, safety rules
│ ├── schemas/ # JSON contracts for runtime artifacts
│ ├── scripts/ # executable gates and BDC runtime
│ └── templates/
├── adapters/
│ ├── claude-code/
│ ├── codex/
│ └── cursor/
├── runtime/ # doctor, dispatch, projection, release checks
├── tests/ # deterministic local checks
├── docs/
│ ├── brand-story.md
│ ├── builder-workflow.md
│ ├── capability-matrix.md
│ └── compatibility.md
└── .github/workflows/smoke.yml