Downshift — AI skill for Claude Code
Deterministic open-source model router — right-sized LLMs per workload.
How to install Downshift
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open tiagovilasboas/downshift and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Downshift does
Deterministic open-source model router — right-sized LLMs per workload. Agent hooks, catalog policy, shadow mode, outcome benchmarks.
Alternatives in AI
- Agentmemory — #1 Persistent memory for AI coding agents based on real-world benchmarks 1.9k ★
- Lumenx Model Onboarding — LumenX 模型接入与文档更新流程 - 厂商文档抓取、model_catalog 更新、运行时范围判断、端到端验证 1.2k ★
- OmniGlyph — Cut your Claude bill 59–70% by rendering bulky LLM context as dense PNG pages — 100% read accuracy, exact per 99 ★
README
Downshift
[](https://github.com/tiagovilasboas/downshift/actions/workflows/ci.yml) [](LICENSE) [](https://go.dev) [](https://github.com/tiagovilasboas/downshift/releases)
**Deterministic model routing for AI workloads.**
Downshift classifies each workload, applies catalog policy, and selects model tier and reasoning effort **before execution**—without an LLM in the routing loop. One Go binary: use it as a **coding harness hook** today; the same core is gateway-shaped for tomorrow.
Subagent / request
↓
classify (rules + optional semantic boost)
↓
policy + catalog → tier & model
↓
execute on the chosen model
**Works today with Claude Code, Cursor, Codex, Antigravity, and KiroCrew.** Single binary, no runtime dependencies, no network calls for routing, no API keys to pick a tier.

**Beta — practical testing phase.** Hooks and adapters are production-minded; harness plans and builds vary in how faithfully they honor `updatedInput.model`. See [Plan compatibility](#plan-compatibility--read-before-installing) before installing.
Why Downshift?
| Right-sized models | Trivial tasks downshift to small tiers; complex work stays on frontier. |
| Deterministic | Scored signals, not an LLM classifier, on the hot path. |
| Testable | downshift try, outcome benchmarks, optional shadow classifier (DOWNSHIFT_SHADOW_WEIGHTS). |
| Observable | Local telemetry (events.jsonl), feedback with explicit --required-tier for training labels. |
| Harness-agnostic core | Adapters only encode I/O; routing lives in `i |
Related Skills
AI To Agent Custom Skill
AI to Agent 上位決策治理 Skill — Context over Dogma, Problems over Benchmarks, People over Models, Outcome over Comp
Stakeholder Analysis Skill For Developing Partnership Strategy
Free AI skill that turns stakeholder analysis into a partnership strategy: stakeholder mapping matrix (impact
Promptforger
A Claude Skill that turns raw, vague "caveman" project ideas into clear, right-sized prompts for another AI or
Effortlane
Effortlane: open-source model and reasoning-effort routing for coding agents. Native Codex auth, Shadow evalua
Evolve Router
A local-first LLM routing gateway that thinks before it routes — a decision model (Jev / laya) picks the most
Product Mode
The most expensive bug in AI-assisted building is shipping the wrong thing, well. A CLAUDE.md for PM + eng tea
Related Agents
GCP Architect
Google Cloud solutions architect: Well-Architected Framework reviews, resource hierarchy and organization poli
Copy Compressor
Compresses source material into infographic-sized copy while preserving facts, voice, hooks, slot clarity, and
Dst
Deterministic Simulation Testing methodology, harnesses, fault injection, and shadow state