Tsubasa — AI skill for Claude Code
Your engineering captain: a repo-resident AI persona with a knowledge graph stored in git.
How to install Tsubasa
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open ramarahmanda/tsubasa and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Tsubasa does
Your engineering captain: a repo-resident AI persona with a knowledge graph stored in git. Agents cite incidents, ADRs and PRs instead of guessing — 47% fewer wrong answers, 40% cheaper on a public benchmark. Claude Code plugin + CLI.
Alternatives in AI
- Pilotfish — Multi-model orchestration layer for Claude Code — the frontier model plans, cheaper models execute, verificati 688 ★
- Llmman — Run any agent on any model, models stored as OCI images 481 ★
- Product Mode — The most expensive bug in AI-assisted building is shipping the wrong thing, well 46 ★
README
tsubasa
**Your engineering captain.** He knows your code, your outages, your history, and your roadmap. He ships you past storms.
[](https://github.com/ramarahmanda/tsubasa) [](https://pypi.org/project/tsubasa/)   
**47% fewer wrong answers · 45% cheaper · 24% faster**
you: why does payment-svc cap its pool at 40?
no captain: pool.go sets it to 40. Likely to avoid exhausting database
connections — a common pattern. You could raise it if your
database handles more concurrent connections, though you may
want to load-test first and monitor p99 during the rollout.
captain: The July 3 flash-sale incident: pool exhaustion took p99 to 8s.
adr-payments-pool-sizing fixed the cap because RDS connections
were the real limit. PR-1841, services/payment/pool.go:88.
goal-replace-apigee retires the layer in front of this service.
Don't invest in the current routing.
Numbers
The measurement is a real agent doing real work: a headless Claude Code session answering questions about four pinned public repos — cloudnative-pg, etcd, kubernetes/enhancements, postgres — scored on what it cites. 72 questions, the same agent with and without a captain, 144 runs, Claude Sonnet 5 in both arms and as the judge.
| vs no-captain baseline | wrong answers | cost | time | tool calls |
|---|---|---|---|---|
| captain | −47% | −40% | −24% | −49% |
| per quest
Related Skills
Bearing
An intel layer for AI coding agents that assumes its own rules might be wrong — and keeps the evidence. Author
Model Identity Skill
Model-agnostic system prompt that aims to keep LLMs honest about their identity, capabilities, and knowledge c
Editorzero
Open-source, self-hostable AI-native docs platform. Humans and AI agents as peer co-editors. Phase 1: architec
Grill With Docs
Grilling session that challenges your plan against the existing domain model, sharpens terminology, and update
1Helm
Self-hosted home for durable AI employees: one resident, one private computer, compounding memory and skills,
Rolepod Brain
Persistent memory for AI coding agents. One binary, nothing resident, no API keys.
Related Agents
Part 19: Repowise — Codebase Intelligence For Coding Agents
60% fewer tokens. 4x faster. Your coding agents stop guessing and start knowing. ---
Autopilot Implementer
TDD implementer for autopilot's fast/cost-efficient mode. Implements ONE PLAN.md work package test-first, runs
Introspect Dont Guess
Resolve REAL IRIS class / table / column / config names instead of guessing. Use before writing SQL or referen