Wingman is a Claude Code plugin that gives non-technical founders a full AI SDLC — an AI Boardroom of agents that plans, builds, secures, and ships production-grade software end to end, with plain-language checkpoints instead of code review.
At a glance:
| Plugin surface | Commands, skills, and 8 fixed Boardroom seats — run node scripts/wingman-health.mjs for the live command/skill counts |
| Eval coverage | Behavioral eval cases across verified/provisional trust levels — run node scripts/wingman-health.mjs for the live count and split |
| Current release | See CHANGELOG.md for the current version and history |
| Install target | Claude Code marketplace + plugin (.claude-plugin/marketplace.json) |
| Runtime dependencies | None — markdown + one dependency-free Node hook script |
/plugin marketplace add LabLaunchPad/Wingman
/plugin install wingman
Then run /wingman:discovery to start a new project, or /wingman:boardroom to get a plain-language review of work already in progress. Every wingman:* command is listed in plugins/wingman/.claude-plugin/plugin.json.
Wingman is built as a Claude Code plugin, and most of the Boardroom's execution mechanism (interactive
founder questions, the plan-approval gate, parallel multi-seat dispatch) is genuinely coupled to
Claude Code's own tool surface — see docs/status/ARCHITECTURE.md §8a/§8b for the
honest account of what is and isn't portable. Two harnesses have a real, scoped starting point:
- Codex CLI and OpenCode — Boardroom seat personas translated into each harness's native
agent format, plus install steps:
plugins/wingman/references/harness-adapters/. Every artifact there is labeledbuilt + tested,authored, unverified, ornot attempted, documented why— nothing is overclaimed as working until it's actually been run against a live install of that harness. - Any harness at all (including a human running plain
git) — the git-push safety gate (dod-structural-gate.mjs's threat-register/traceability/test checks) has a harness-agnostic fallback:plugins/wingman/scripts/install-git-hooks.mjswires it up as a real.git/hooks/pre-pushhook, with zero per-harness adaptation. This is the one piece confirmed working outside Claude Code.
Instead of asking a founder to read code or a diff, Wingman gates every stage of the SDLC through a Boardroom checkpoint: 7 C-suite-style seats plus Design (CEO / CPO / CMO / CTO / CISO / CFO / Research / Design) examine the plan or change in parallel and hand back one short, jargon-free go/no-go summary, consolidated into a grouped Business / Technical / Finance / Research report. The founder makes the call; Wingman never assumes silence means approval.
flowchart LR
subgraph P1["1. Problem Definition"]
direction LR
A1[discovery] --> A2[research-synthesis] --> A3[personas-jobs]
end
subgraph P2["2. Logic & Mapping"]
direction LR
B1[journey-mapping] --> B2[define]
end
subgraph P3["3. Lean Design"]
direction LR
C1[information-architecture] --> C2[uxflow] --> C3[wireframes] --> C4[visual-design-system]
end
subgraph P4["4. Architecture & Build"]
direction LR
D1[prototype-usability] --> D2[architecture] --> D3[implementation-planning] --> D4[build]
end
subgraph P6["6. Launch & Iterate"]
direction LR
E1[ship]
end
P1 --> P2 --> P3 --> P4 --> P6
P4 -.->|every one of stages 1-12 has its own solo checkpoint| CP{{Boardroom}}
14 named stages, 14 founder-visible checkpoints — every stage 1 through 12 records its own solo Boardroom checkpoint immediately after itself (no bundling), plus build's own Definition-of-Done gate (folding in the security pass) and ship's. See docs/status/ARCHITECTURE.md §4d for the full stage list, traceability prefixes, and the founder-approved reversal of an earlier, more-bundled 7-stage/3-checkpoint model. A change-triage skill runs at intake (discovery, build, hotfix) to route a genuinely small, non-trust-boundary fix to a lighter path instead of the full sequence — see §4d's note on this deliberate, explicitly-recorded partial exception.
The agent population is deliberately lazy, not exhaustive:
- A fixed 7+1-seat Boardroom is always present — it only renders verdicts, never writes code.
- A small set of department leads (Product, Engineering, QA always active; Design, Data, Legal/Security, DevOps, Growth created only when a project's real complexity calls for them) do the actual build-time work.
- A Management Board of coordinators activates only once a project crosses 3+ active conditionally-created department leads.
- Narrow specialists (a 56-role candidate catalog) are promoted one at a time via
/wingman:evolve, only after proven, repeated need — never bulk-created.
See docs/status/ARCHITECTURE.md for the full model and docs/roadmap/AGENT-ROSTER.md for the complete specialist catalog.
The pipeline is built and behaviorally tested, not just scaffolded:
- Commands — 14 named SDLC pipeline stages (
discovery/research-synthesis/personas-jobs/journey-mapping/define/information-architecture/uxflow/wireframes/visual-design-system/prototype-usability/architecture/implementation-planning/build/ship) plus adaptive commands (audit,boardroom,launch,hotfix,harness,telemetry,retro,learn,evolve,over-engineering-review,bloat-audit,debt-ledger,research,advisory,incident,dogfood,knowledge-export,post-launch,review,test). Runnode scripts/wingman-health.mjsfor the live, exact count. - Skills covering discipline (
engineering-minimalism,verification-before-completion), mechanics (git-pr-workflow,security-checklist), and adaptive output (visual-founder-output,plain-language-checkpoint). - 8 fixed Boardroom seats (7 C-suite-style + Design), dispatched in parallel and never writing code.
evals/ holds a lightweight behavioral eval harness (not just structural validation): eval cases in a mix of verified (passed 2+ differently-shaped scenarios including a negative case) and provisional (passed one real run) — see evals/README.md for the trust-level bar. Covers every command and the high-value skills, including full end-to-end pipeline runs against realistic projects (one adversarial run producing a real DO NOT SHIP). Run node scripts/wingman-health.mjs for a live, read-only snapshot of these numbers straight from the repo — it's the source of truth this table is generated from, not a number to trust in prose.
See docs/status/PROJECT.md for exact build/eval status and decisions log, and docs/status/ARCHITECTURE.md for what's built versus deliberately deferred. Most of this was verified in a sandboxed testing environment; real dogfooding passes (actual claude CLI install, live pipeline runs via /wingman:dogfood) have also happened and found real gaps, now fixed — see docs/HUMAN-TODOS.md for what real dogfooding still needs.
This README is written to be skimmed top-to-bottom in under a minute. If you're an AI coding agent working in this repo instead of a human reading it, start with AGENTS.md — the canonical project brief, following the open AGENTS.md standard (CLAUDE.md is a symlink to it, kept for tools that look for that filename specifically) — rather than re-deriving conventions from source. See docs/status/ARCHITECTURE.md §8a for the honest scope of what's portable to a non-Claude-Code harness today (two skills) versus what isn't (the rest).
docs/status/ARCHITECTURE.md— the hybrid Boardroom/department/specialist model and why it's shaped this way.docs/status/GOVERNANCE.md— a one-page index of where org governance, policy enforcement, and benchmarks/metrics each already live in this repo.docs/roadmap/AGENT-ROSTER.md— the full specialist candidate catalog, organized by department.docs/status/PROJECT.md— current build/eval status, decisions log, and roadmap.CHANGELOG.md— release history.docs/HUMAN-TODOS.md— what's blocked on a human rather than more engineering (publishing, demo content — seedocs/DEMO-CHECKLIST.md).evals/README.md— how the behavioral eval harness works and what's been verified.ATTRIBUTIONS.md— provenance for design patterns adapted from vendored reference repositories.SECURITY.md— vulnerability disclosure and this repo's actual trust boundaries.CONTRIBUTING.md— how to propose a change.