A persistent project assistant for Claude Code. Bootstraps new projects, softly adopts existing ones, and stays as a quiet helper — maintaining the wiki, recommending the right model per task, surfacing past decisions, and tracking how the project evolves.

Built for vibe-coders first: people who describe ideas in plain language and want the right defaults picked for them. Works for technically literate users too — every decision is optional, and you can switch off anything that gets in the way.

# 1. Install the skill in your project
npx skills add dtdesigner36/jarvis-starter --yes

# 2a. New project? Start fresh:
> jarvis start: Telegram bot for reminders with a Next.js admin panel

# 2b. Existing project? Adopt without breaking anything:
> jarvis adopt

# 3. That's it. JARVIS lives in .jarvis/ and runs via hooks.
#    Need something? Ask it:
> jarvis status
> jarvis route "refactor the auth module"
> jarvis find "pdf parser"
> jarvis audit

Starting a Claude Code project means re-inventing the same scaffolding every time: CLAUDE.md rules, hooks, wiki, which skills to install, which model to pick for which task. And once the project grows, keeping docs and decisions in sync with the code is a losing battle.

JARVIS solves both:

• On start — sets up architecture for the project type (Telegram bot, web app, API, landing, game, parser, mobile, library, CLI, LLM-agent, and more) using archetype-specific overlays.
• On adopt — for existing projects, runs a gap analysis and installs only the features you're missing. Adds a single marker line to CLAUDE.md; never overwrites your existing rules. Doesn't replace your hooks — wraps legacy hook names with a sentinel-guarded block or installs new jarvis-*.sh files alongside. Everything new lives in .jarvis/ or in separate jarvis-*.sh hook files.
• Ongoing — passive hooks maintain the wiki, recommend the right model (Opus / Sonnet / Haiku) and plan-mode for each task, resurface already-solved problems, and track project evolution. 0 tokens until a hook fires — each core hook runs once per tool-use or user turn, then returns silently when there's nothing to say.

> jarvis start: <description of what you want to build>

JARVIS walks you through:

1. Requirements clarification
2. Stack selection
3. Relevant skill picks from the curated registry, ranked by archetype + stack overlap (GitHub-query discovery is model-guided — the install step is yours)
4. Infrastructure rollout (CLAUDE.md, hooks, wiki/)
5. Initial optimization advice

> jarvis adopt

Soft integration via gap analysis. JARVIS:

• Reads your project (stack, CLAUDE.md, hooks, docs, secret-scanners, issue tracking)
• Detects what you already have (living docs, husky pre-commit, linear integration, etc.)
• Shows a gap matrix — proposes only the features you're missing
• Installs those features in .jarvis/ namespace + new jarvis-*.sh hook files
• Adds one marker line to your CLAUDE.md (<!-- jarvis-starter-adopt: see .jarvis/state.md -->); your own rules are never rewritten
• Leaves your existing hooks and docs in place — pre-existing legacy hook files are wrapped in a sentinel-guarded block you can undo with safe-uninstall.sh, and docs/ is never overwritten

If you run jarvis start in an existing project, JARVIS detects the dev-stage (git history, mature lockfile, existing CLAUDE.md, living docs, CI) and auto-switches to Adopt. Force full bootstrap with jarvis start --force if you really mean it.

JARVIS commands come in two flavours. The executable ones run shell scripts directly. The model-prompted ones are markdown instruction files that Claude reads on demand — they guide Claude through a workflow, they are not standalone programs.

Executable shell commands:

Model-prompted workflows (markdown Claude reads when invoked):

> jarvis school on

Creates a school-wiki/ with a topic index for your stack. Each topic is a short stub — what it is, where it lives in this project. When you want to go deep:

> jarvis school topic prisma-migrations

JARVIS: Let's walk through migrations. In your project:
  - 3 migrations (latest one 5 days ago)
  - Schema has X models

  What do you already know about migrations in general? Have you ever
  dropped a column from an existing table?
  ...

Turn off with jarvis school off. JARVIS never proactively suggests school-mode — it's your call.

jarvis-starter/
├── SKILL.md                  # Entry point
├── README.md                 # This file
├── NOTICE.md                 # Attributions
│
├── core/                     # Always-on hooks (0 tokens until a hook fires)
│   ├── wiki-maintenance/     # Watches code → proposes wiki updates
│   ├── task-routing/         # Classifies prompts → recommends model
│   ├── memory-recall/        # Surfaces already-resolved topics
│   ├── focus-tracker/        # Passively tracks current focus
│   └── security-watch/       # Detects hardcoded secrets and .env leaks
│
├── on-demand/                # Markdown instructions Claude reads on demand
│   ├── skill-discovery/      # jarvis find — registry match + GitHub query hints
│   ├── security/             # jarvis security — audit workflow
│   ├── evolve.md             # jarvis evolve
│   ├── decide.md             # jarvis decide
│   ├── suggest.md            # jarvis suggest
│   ├── docs.md               # jarvis docs
│   ├── audit.md              # jarvis audit
│   └── self-audit/           # jarvis self-audit — real shell script (scripts + report.sh)
│
├── plugins/                  # Optional extensions (off by default)
│   └── school-mode/
│
└── archive/                  # Rarely needed after bootstrap
    ├── bootstrap/            # Phase 0-6 bootstrap + brownfield-adopt
    ├── archetypes/           # Tier 1 (10 full) + Tier 2 (19 descriptions)
    ├── templates/            # Universal + archetype overlays
    └── subcommands-rare/     # remember, forget, history, focus, optimize

1. Quality > optimization — but both matter
2. Quiet butler — silent until called
3. Plain language — for vibe-coders
4. Wiki as first-class — always current, via hooks
5. Respect what's already there — brownfield-safe by default (Adopt mode)
6. Extensible — plugins for different work modes

• Planned features and current priorities → GitHub Issues and Releases
• Full changelog → CHANGELOG.md

Issues, feature suggestions, and PRs welcome. See CONTRIBUTING.md for how to file a good bug report or propose a feature. Discussion happens in GitHub Discussions.

If you've built a related Claude Code skill, open an issue — happy to add it to the known-skills registry referenced by jarvis find.

JARVIS-Starter stands on the shoulders of others:

• @alinaqi — orchestration patterns, Stop Hooks approach, and persistent memory concepts in claude-bootstrap
• @pbakaus — Context Gathering Protocol, .impeccable.md persistence pattern, design philosophy, and design-skill templates. The foundation for JARVIS's "context first" approach
• Emil Kowalski — UI polish philosophy and the "invisible details" design engineering mindset
• @leonxlnx — agency-level design direction and the visual taste framework
• Anthropic — skill-creator patterns (progressive disclosure, description optimization, subagent testing loops)
• @wcpaxx — brownfield detection approach and the "analyze before asking" idea
• @travisvn/awesome-claude-skills — curated community registry as a reference for our skill-discovery

Each solved a piece of the puzzle of embedding Claude into a workflow. JARVIS combines these ideas and adds its own angle: a persistent assistant across the project's lifetime, not a one-shot bootstrap.

MIT — see LICENSE and NOTICE.md for formal attributions.