discipline
AI coding agent workflow harness
Ship clean code. Not AI bloat.
The right level of engineering discipline for every task. Reuse what exists, match the project, and go playoff mode only when the stakes demand it.
discipline
AI writes fast.
It should also know
when to stop.
+ src/components/LogoutButton.tsx
+ src/styles/logout-button.css
+ src/theme/logout-tokens.ts
+ src/services/auth-controller.ts
+ src/hooks/useLogoutFlow.ts
<Button variant="default"
onClick={handleLogout}>
Logout
</Button>
- 01 Existing code
- 02 Standard library
- 03 Native feature
- 04 Installed dependency
- 05 Smallest new implementation
Every task
is not game seven.
Start light. Add process only when risk earns it. mancode lets the same coding agent move from daily practice to production-grade review without changing tools. Newer models that already self-review get a hard stopping point; models that need guidance keep an explicit quality gate.
solo
Move fast.
Stay on system.
For everyday changes that need context, not ceremony—even with newer models that already self-review.
- Lightweight hooks
- One bounded diff self-check
- No extra reviewer loop
/manba
Reproduce.
Diagnose. Verify.
For defects that need a root cause and a real regression check.
- Reproduce the failure
- Drive real user paths
- Verify the fix in runtime
/man
Plan every
possession.
Align requirements, compare options, and approve a durable plan before choosing delivery depth.
- Guided requirement alignment
- Lightweight solo handoff
- Optional full nine-step governance
/manteam
Five on floor.
One memory.
For teams that need shared decisions to survive across agent sessions.
- Shared project memory
- Decision records
- Conventional commits
/manps
Find the rust
before tipoff.
For maintenance passes that surface debt without rewriting the project.
- Stale TODO detection
- Dependency checks
- Markdown + JSON reports
/man workflow
From scout report
to final buzzer.
Every planning run leaves a readable trail under
.mancode/<namespace>/workflows/<ULID>/. After approval, hand the plan to lightweight solo
delivery or continue through the full governed workflow.
- 01Scout reportMap code, risks, unknowns
- 02ClarifyResolve the real requirement
- 03PlanCreate verifiable steps
- 04Plan gateSolo, full governance, plan-only, or revise
- 05ImplementHead Coach runs the play
- 06VerifyBuild, lint, test, smoke
- 07Targeted reviewEvidence-backed quality findings
- 08Risk defenseSecurity review only when risk requires it
- 09Close outOne remediation, verify, stop
Read the room
before writing.
mancode detects the project shape, available validation commands, existing components, CSS variables, and design signals. UI guidance only activates when the project actually has a UI.
your-project/
├── .mancode/
│ ├── shared/context/project.json
│ ├── local/cache/style-tokens.json
│ ├── local/workflows/<ULID>/
│ └── shared/team/
├── .claude/
├── .cursor/rules/
├── .agents/skills/
└── AGENTS.md
✓ context stays with the code
✓ no telemetry sent
One playbook.
Your tools.
mancode is a workflow layer, not another coding agent. Claude Code gets the richest native integration; other tools receive durable rules, skills, or instruction files.
One task.
Another session.
mancode keeps goals, requirements, plans, validation results, and handoff notes under a stable TaskRef. A new chat or CLI session can resume that task and load only the Context Pack it needs, without depending on the previous window staying open.
mancode context resume <TaskRef> --session <id> --client <client>Explicit identity
Ready at the buzzer
Put mancode
in the rotation.
No account, dashboard, API key, or new editor. Initialize it inside the project you already use—Windows CMD, PowerShell, Git Bash, macOS, and Linux are supported.
Read the docs ↗npm install -g mancode
cd your-project
mancode init
Less ceremony for the easy work.
More discipline for the work that matters.