Doby Cuts Claude Code Navigation Tokens by 95% with Spec-First Workflow

A spec-first fix workflow that slashes navigation tokens 95% and enforces plan docs as source of truth before code changes.

What Changed — The specific update

Doby is a new open-source skill for Claude Code that rethinks how you navigate and modify codebases. Instead of letting Claude burn 2,000-5,000 tokens just figuring out where to make changes, Doby pre-compiles an index of your project's keywords, plan docs, code files, symbols, and status into a flat-line index that costs ~100 tokens per lookup.

Install it with:

git clone https://github.com/changmyoungkim/doby ~/.claude/skills/doby
cd your-project
bash ~/.claude/skills/doby/install.sh
# Then in Claude Code:
/doby build

What It Means For You — Concrete impact on daily Claude Code usage

The real innovation isn't just the token savings — it's the enforced workflow. Doby implements a 7-phase process that forces spec-first development:

1. Phase 0: doby resolve — 2 grep calls to find docs + code (~100 tokens)
2. Phase 1-2: Spec Review → Update — Update spec first, get user approval
3. Phase 3: Plan Verification — Feasibility, architecture, risk checks (loop until consensus)
4. Phase 4: Implementation — Parallel Haiku agents, Opus verifies consistency
5. Phase 5: Integration Test — Run tests, verify
6. Phase 6: doby sync — Update INDEX + consistency check
7. Phase 7: Complete — Report + log

This eliminates the biggest hidden cost in LLM-driven development: rework cycles. Instead of writing code, testing, finding it violates the spec, and iterating 2-4 times, Doby verifies the plan before any code is written.

Try It Now — Commands, config, or prompts to take advantage of this

Configure .dobyrc.json in your project root:

{
  "scan_dirs": ["src", "backend", "frontend"],
  "file_extensions": [".py", ".ts", ".tsx", ".js"],
  "plans_dir": ".omc/plans",
  "wiki_dir": ".omc/wiki",
  "exact_file_map": {
    "src/index.ts": "app-entry"
  },
  "directory_rules": {
    "src/auth/": "auth-system",
    "backend/services/payment/": "payment-processing"
  },
  "keyword_to_doc": {
    "auth": "auth-system",
    "cache": "caching-layer"
  }
}

Matching priority: exact_file_map > directory_rules > keyword_to_doc > fuzzy match.

The index is layered — 90% of daily work stays in L1 (flat-line index, ~100 tokens). L2 (wiki pages, ~500 tokens) triggers on "why?" questions only. L3 (semantic RAG, ~300 tokens) activates on L1 misses. L4 (auto-compile, ~2,000 tokens) is manual trigger only. Core savings are never diluted.

gentic.news Analysis

Doby follows the trend we've seen with other Claude Code skill enhancements — developers are increasingly treating token budgets as the primary optimization target. This aligns with our previous coverage of CLAUDE.md best practices where structured context reduces waste.

The spec-first approach is particularly notable because it addresses a systemic failure in LLM-assisted coding: the tendency to generate code that looks right but violates implicit design constraints. By forcing plan verification before implementation, Doby mirrors what disciplined human teams do in code review — but automates it.

For Claude Code users, this is a significant workflow upgrade. The 90-95% token reduction on code changes means you can fit more iterations into a single context window, reducing the need for expensive context refreshes. If you're hitting context limits on large refactors, Doby's layered index approach is worth adopting immediately.

---

Originally published on gentic.news