End-to-end autonomous software design, implementation, and adversarial review in a single pipeline
- Admiral orchestration — one front door for design, build, review, investigation, release, resume, and skill/team creation across 47 skills.
- Save protocol — persisted runs, resumable checkpoints, audit trails, and
recoverable handoffs in
skillset-saves/. - Runtime harness — hook-based routing reinforcement, guarded writes, dangerous-command checks, and failure-trajectory recovery hints.
- Gated pipelines — design, build, review, browser automation, release, safety, testing, and QA workflows with clear ownership boundaries.
- Adversarial review — gatekeepers, bug/security/quality/frontier lenses, cso and mr-robot pressure testing, and fail-loud deterministic checks.
- Templates and doctrine — grill-me intake, shadcn/ui component templates, API contracts, handoff templates, and reusable skill-maker packaging patterns.
Without Supreme Team, asking an AI assistant to "build me an app" produces a single-pass attempt with no structured validation, no adversarial review, and no way to resume if the conversation ends mid-task.
With Supreme Team, the same request enters through one front door (admiral) and flows through a phased pipeline where each deliverable is challenged by adversarial gatekeepers before the next phase begins. Design specs are validated before code is written. Code is security-audited before review. Review findings are evidence-checked before delivery. Every artifact is saved to disk for cross-session resume and audit, and a runtime harness deterministically blocks dangerous actions and steers lifecycle work through the orchestrator.
- One enforced entry point — admiral is the primary entry orchestrator; tell it what you want and it routes through the right phases under one intake, one persisted run, and one cross-stage gate
- Adversarial quality gates — four gatekeepers challenge every deliverable, backed by a deterministic gate engine; approval is earned, never assumed
- Runtime harness — stdlib hooks block dangerous commands and guarded-path writes, detect degenerate trajectories, and reinforce entry routing
- Cross-session persistence — pipeline state, deliverables, and audit trails are saved to disk so you can resume exactly where you left off
- Flexible execution — run the full pipeline, any subset of phases, individual skills, or the standalone tool groups
- Skill & team creation — skill-maker drafts, evals, reviews, and packages new Claude skills and coordinated skill teams
- No platform lock-in — plain Markdown files that work with any AI tool that reads skill definitions
Supreme Team contains 47 skills: a three-stage delivery pipeline managed by sub-orchestrators under the top-level admiral, three cross-cutting Admiral-pipeline components, and four standalone tool groups.
| Group | Orchestrator / Tier | Skills | Purpose |
|---|---|---|---|
| Admiral Layer | admiral | 2 | Top-level orchestration + cross-stage validation |
| Design | commander | 6 | Requirements, planning, architecture, API contracts, design system, impl spec |
| Build | build-management | 8 | Implementation, testing, security hardening, completeness, debugging, health |
| Review | code-chief | 11 | Bug detection, code quality, security, security leadership, pen-testing, frontend, visual QA, DX |
| Investigation | — (in-scope) | 1 | Disciplined root-cause analysis |
| Skill Maker | skill-maker | 3 | Create, review, and package Claude skills and teams |
| Session Memory | — (in-scope) | 1 | Cross-session checkpoints and accumulated learnings |
| Browser Automation | standalone | 4 | Launch, drive, authenticate, and share a browser session |
| Release & Deployment | standalone | 4 | Release orchestration, merge-and-deploy, deploy config, release notes |
| Safety Guardrails | standalone | 4 | Intent checks and write-boundary locks (guard / careful / freeze / unfreeze) |
| Testing & QA | standalone | 3 | Test-and-fix QA, report-only QA, performance benchmarking |
Every pipeline sub-orchestrator delegates to its specialists in sequence and validates each phase through its own gatekeeper before advancing.
See docs/skills.md for the complete inventory and docs/architecture.md for the pipeline flow.
admiral is the single front door for the entire delivery lifecycle. The
entry-routing doctrine fixes the gap that descriptions alone cannot close: a
request like "design this system" or "investigate this bug" is initiated through
admiral so the run gets one intake, one persisted state, and one cross-stage
gate. In-scope skills reached cold hand off to admiral first; standalone tools
(Tier 4) run directly. A UserPromptSubmit hook reinforces this deterministically
once registered.
See docs/routing.md for the tiers, precedence, and loop guard.
Supreme Team enforces quality through four adversarial gatekeepers at two
levels. Per-phase gatekeepers (gatekeeper-design, gatekeeper-build,
gatekeeper-code) validate work within their sub-pipeline. The cross-stage
gatekeeper (gatekeeper-admiral) validates at the boundaries between stages. A
fifth adversarial gate, skill-reviewer, scores skills inside the skill-maker
pipeline. Every gatekeeper is backed by a shared deterministic gate engine that
turns mechanically checkable conditions into a fail-loud pass.
Every gatekeeper verdict is one of:
- APPROVED — advance to the next phase
- REVISE — return with specific findings to address (max 2 cycles)
- ESCALATE — surface the blocking issue to the user
A review that finds nothing is treated as the most suspicious review of all.
See docs/gatekeepers.md for the full pattern.
A stdlib-only runtime harness deterministically enforces the parts of the
contract that can be checked mechanically: pre_tool_use.py blocks dangerous
commands and writes into a frozen/guarded boundary, post_tool_use.py detects
degenerate trajectories and injects a recovery hint, and user_prompt_submit.py
reinforces entry routing. The hooks fail open; the gatekeeper gate engine
fails loud. Hook registration is explicit opt-in and uses each host's native
configuration or plugin mechanism.
See docs/harness.md and skills/harness-doctrine.md.
Pipeline state is automatically saved to skillset-saves/ in your project
workspace as the pipeline runs, providing cross-session resume, crash recovery
(lease-based locking, idempotent gatekeeper submissions), a chronological audit
trail, deliverable backup, and graceful degradation if saves fail.
See docs/persistent-saves.md for details.
- Context window dependent — large projects can exceed an AI assistant's context window. The save system mitigates this with reference-mode tiers, but very large codebases may still require manual chunking.
- LLM accuracy — the pipeline enforces structure and adversarial review, but output quality is bounded by the underlying model. Gatekeepers catch many issues but are not infallible.
- No runtime execution — Supreme Team generates artifacts (code, configs, runbooks) and orchestrates release flows, but a human or CI system runs the actual deployment commands the release tools produce.
- Hook enforcement is host-dependent — deterministic entry routing and action guards require the runtime harness hooks to be registered for the active host. Without them, those layers fall back to advisory doctrine.
- Single-session concurrency — the lease-based lock system is advisory. Running two sessions against the same project simultaneously can cause conflicts.
See QUICK-START.md for installation steps and first-use instructions.
| Document | Description |
|---|---|
| QUICK-START.md | Installation and first-use guide |
| Install.md | Detailed installation procedure (AI-agent and manual) |
| AGENTS.md | Authoritative skill manifest for tool discovery |
| docs/architecture.md | Pipeline architecture, flow diagrams, and execution modes |
| docs/skills.md | Complete skill inventory |
| docs/routing.md | Entry-routing doctrine and skill tiers |
| docs/gatekeepers.md | Gatekeeper pattern and the deterministic gate engine |
| docs/harness.md | Runtime harness: hooks and gate engine |
| docs/persistent-saves.md | Save system, resume, and audit trails |
| docs/direct-invocation.md | Standalone skill usage and routing tiers |
| docs/directory-structure.md | Repository and installed layout reference |
Built by TykoDev · Supreme Team