devjar

Scan your project into a Claude-ready knowledge map. Rewrite vague prompts into structured ones. Use 60–70% fewer tokens on any AI coding tool.

Every Claude Code session starts blind. You re-explain the same context, re-read the same files, burn tokens on clarification. devjar fixes both.

Quick start

npm install -g devjar
cd your-project
devjar init

Your project is now mapped. Claude reads CLAUDE.md at session start and never re-reads your files again.

For prompt normalization, pick a provider once:

ollama pull llama3.2          # free, runs locally
devjar config --provider ollama --model llama3.2
devjar prompt "fix login bug"

Commands

| Command | What it does | |---|---| | devjar init | Scan project → generate CLAUDE.md (stack, file map, rules) | | devjar prompt "text" | Rewrite vague input into STAR-C format, copy to clipboard | | devjar update | Regenerate CLAUDE.md after major changes | | devjar update --watch | Live watcher — updates file map on every save | | devjar stats | Token savings estimate, streak, per-project breakdown | | devjar config | Set AI provider, or run setup wizard |

Provider setup

Ollama — free, local, no account needed (recommended)

# Install from https://ollama.com, then:
ollama pull llama3.2
devjar config --provider ollama --model llama3.2

Cloud providers

devjar config --provider gemini    --key AIza...    # free tier at aistudio.google.com
devjar config --provider anthropic --key sk-ant-...
devjar config --provider openai    --key sk-...

First time you run devjar prompt with no config, the setup wizard runs automatically.

What it looks like

$ devjar init
⬡ devjar init → /your-project
────────────────────────────────────────────────────
Phase 1  Scanning project structure...
  ✓ 115 files | Stack: React + Vite
Phase 2  Pattern extraction via Haiku...
  ✓ 5 rules extracted from 4 key files
Phase 3  Writing CLAUDE.md...
────────────────────────────────────────────────────
✓ CLAUDE.md generated — 115 files scanned, 5 rules extracted
  → /your-project/CLAUDE.md

$ devjar prompt "login not working after deploy"
⬡ devjar prompt
────────────────────────────────────────────────────
  Provider: ollama/llama3.2
  Input:    login not working after deploy
  ✓ CLAUDE.md loaded (7236 chars)

─── STAR-C Prompt ───────────────────────────────────
S — care-home — React + Vite, auth via httpOnly cookie
T — Debug login failure introduced after deployment
A — Check src/routes/auth.js COOKIE_OPTIONS (SameSite attr)
    Verify KV session creation on login success
R — Login succeeds, sc_token cookie set, /api/auth/me returns user
C — Don't change PBKDF2 iterations (CF Workers hard cap: 100k)
    Don't rotate JWT_SECRET without flushing KV sessions
────────────────────────────────────────────────────

Copy to clipboard? (Y/n) y
✓ Copied

How it works

devjar init walks your project, reads package.json + config files, then sends 3–5 key files to your AI provider and extracts naming conventions, patterns, and 5 unbreakable rules. Writes everything to CLAUDE.md.

devjar prompt sends your vague input + your CLAUDE.md to your provider. Gets back a structured prompt with the exact files to check, expected result, and constraints inferred from your codebase. One provider call ≈ $0.00 with Ollama.

Requirements

• Node.js 18+
• devjar init — fully offline, no provider needed
• devjar prompt — Ollama (free) or any cloud API key
• devjar stats / devjar update — fully offline

License

MIT