armory

armory

Curated, production-grade skills, agents, hooks, rules, commands, utilities, and presets for AI coding agents. No magic, no demos — battle-tested workflows built for developers who use AI seriously.

---

Overview

armory is a collection of packages for Claude Code and Claude.ai. Each package is a self-contained prompt or automation unit that extends Claude with a repeatable, opinionated workflow for a specific task domain. Packages span seven types: skills, agents, hooks, rules, commands, utilities, and presets.

Philosophy: Packages in this collection are practical and context-free. They define the how, not just the what — covering inputs, outputs, edge cases, and failure modes. They are tested in real workloads, not constructed as examples.

Intended for developers who treat AI coding agents as a serious part of their workflow.

Planning Lifecycle

Use the planning packages in sequence, not as interchangeable formats:

unknown destination → decision-map → known design → plan-prompts → task-decomposer / plan-review → milestone-runner → stacked-prs

Situation	Package	Output
Destination or design is unresolved	decision-map	Tracker-backed decision map, frontier, and recorded decisions
Design is known and needs milestones	plan-prompts	DEVELOPMENT_PLAN.md and EXECUTION_PROMPTS.md
Feature is specified and needs implementation slices	task-decomposer	Phased task board with dependencies and test matrix
Existing plan needs a pre-implementation challenge	plan-review	Severity-ranked plan review
Approved plan needs execution across sessions	milestone-runner	Reviewed, merged milestone PR stack

decision-map stops at resolved decisions. It recommends plan-prompts after a complete map and never creates implementation milestones or tasks.

---

Package Catalog

Agents — Orchestrators

Orchestrator agents compose skills and other agents into multi-phase workflows. Each can run solo or be spawned by another agent via the Agent tool.

Agent	Model	Description
team-lead	opus	Meta-orchestrator — decomposes multi-domain requests, delegates to specialized agents, synthesizes results
codebase-auditor	sonnet	Unified quality assessment — spawns code-reviewer, security-reviewer, secret-scanner in parallel, merges report
project-architect	opus	Phased requirements discovery producing architecture documents with diagrams and tech stack justification
project-planner	sonnet	Task decomposition with dependency mapping, three-point estimates, milestone timelines, and risk logs
research-analyst	opus	Multi-source investigation with parallel agents across web, academic, video, and competitive sources
idea-scout	opus	Business idea validation — Lean Canvas, parallel market/competitive/feasibility research, weighted scorecard
full-stack-builder	opus	End-to-end implementation from spec — scaffolding, sprints, quality passes, documentation, pre-delivery review
release-captain	sonnet	Ship lifecycle with quality gates — pre-flight, secret scan, changelog, version bump, PR creation
proposal-writer	opus	Technical proposals with ROI calculations, three-tier pricing, and Problem-Agitate-Solve framing
content-strategist	sonnet	Multi-channel content creation with per-channel adaptation and automated quality passes
media-producer	sonnet	Visual and video format router — selects the right skill based on concept type and output needs
skill-librarian	sonnet	Explicit reflective capture — queues privacy-safe lessons, reports status, and reviews approved package changes

Agents — Analyzers

Agent	Model	Description
code-reviewer	sonnet	Multi-phase code review with severity-ranked findings
security-reviewer	sonnet	OWASP Top 10 vulnerability scanning
secret-scanner	haiku	Pre-commit detection of hardcoded credentials
skill-router	haiku	Outcome-weighted package routing using historical eval results
test-engineer	sonnet	Co-evolutionary skill evolution with generate-verify-refine loops

Model routing: Agents marked opus request Claude Code's opus alias with xhigh effort by default. The declared target matrix records the alias and an expected fully disclosed model-id prefix; a dated snapshot is an observed receipt field, never inferred from this declaration. Use max effort only for genuinely hard novel problems (diminishing returns, overthinking risk); use high for concurrent sessions or cost-sensitive work.

Runtime model-fit target matrix

model_fit.yaml declares the client/model/effort/profile/telemetry matrix. Its current target requests the Claude Code opus alias, requires a disclosed model ID beginning claude-opus-, and requires observed input-token and output-token fields.

Validate a profile without creating a model request:

uv run python scripts/model_fit.py --target claude-code-opus-xhigh --profile core --dry-run
uv run python scripts/model_fit.py --target claude-code-opus-xhigh --profile full --dry-run

Dry runs read claude --version and resolve repository package definitions only. They report discovery metadata and explicitly state that package content was not loaded. They do not install a profile or make a paid call.

Use --live --output /path/to/receipt.json only when a model request is authorized. The probe disables built-in tools, suppresses ambient MCP configuration, and does not persist sessions. Its receipt stores derived fields only: declared target/profile/tool surface, client version, an observed model that must match the target's declared prefix, required token fields, wall time, and terminal metadata. Raw prompts, assistant text, and stream events are never retained. A missing or changed required field exits non-zero; the probe never estimates a token value.

Skills — Development & Tooling

Skill	Description
agent-builder	Build AI agents using the Claude Agent SDK and headless CLI mode — covers tool definitions, MCP servers, and programmatic orchestration
github	GitHub CLI operations via gh — issues, PRs, CI/Actions, releases, search, REST/GraphQL API, with error handling and automation workflows
filesystem	File and directory operations via Claude Code built-in tools — replaces the Filesystem MCP server with native Read, Write, Edit, Glob, Grep
mcp-to-skill	Convert MCP servers into on-demand skills to reduce active context window token usage
gpu-optimizer	GPU optimization for consumer GPUs (8-24GB VRAM) — PyTorch, XGBoost, CuPy/RAPIDS, memory management, and CUDA tuning
tavily	AI-optimized web search and content extraction via Tavily API with structured output parsing
test-harness	Comprehensive pytest suite generation — happy path, edge cases, error conditions, fixtures, mocks, async, parametrized tests
debug-investigator	Systematic debugging framework — hypothesis-driven investigation with bisection, log analysis, instrumentation, and minimal reproduction
project-context-setup	Scaffold repo-local agent context — issue tracker rules, triage labels, domain glossary layout, ADR lookup, agent brief conventions
stacked-prs	Manage dependent branch stacks and stacked pull requests — inspect, split, publish, sync, validate, merge, and clean up stack topology
plan-prompts	Generate adaptive development plans, execution prompts, and an ignored local design-evidence ledger from source docs
decision-map	Map unresolved architecture, policy, and scope decisions before implementation planning begins
milestone-runner	Run milestones through serialized design reconciliation, reviewed PR gates, CI, cleanup, and release preparation
to-markdown	Convert any file or URL to clean Markdown via MarkItDown — PDF, DOCX, XLSX, PPTX, HTML, images, audio, CSV, JSON, XML, YouTube, EPub
web-fetch	Web content fetching via curl and WebFetch — replaces the Fetch MCP server with native HTTP operations and jq parsing
lightpanda-browser	Lightweight headless browser automation via Lightpanda + agent-browser CDP — 9x lower memory, 11x faster, for scraping, DOM extraction, and form automation
skill-library	Agent-native catalog for browsing, installing, updating, syncing, and removing armory skills from within a Claude Code session
env-validator	Validate .env files against project requirements — missing vars, type mismatches, insecure defaults, .env.example drift