agentleFS
dhriyatinandu-botdhriyatinandu-bot/claude-skills/.claude/skills/llm-router/SKILL.md

llm-router

Skill0 starsChanged 9 months agoMITView on GitHub

Selects the optimal LLM model and provider for each task based on complexity, cost budget, and capability requirements. Routes cheap tasks to Haiku/GPT-4o-mini and complex tasks to Sonnet/Opus/o1.

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 4 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Wrong: Only optimizing for cost, ignoring that Opus takes 5-10x longer than Haiku.

    line 164

“Common” is measured: kept by at least 6 in 10 of the 4 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

LLM Router

Selects the optimal LLM model for each task. The single biggest cost lever in multi-agent systems — intelligent routing saves 45-85% while maintaining 95%+ of top-model quality.


When to Use

✅ Use for:

  • Deciding which model to call for a specific task
  • Assigning models to DAG nodes in agent workflows
  • Optimizing LLM API costs across a system
  • Building cascading try-cheap-first patterns

❌ NOT for:

  • Prompt engineering (use prompt-engineer)
  • Model fine-tuning or training
  • Comparing model architectures (academic research)

Routing Decision Tree

flowchart TD
  A{Task type?} -->|Classify / validate / format / extract| T1["Tier 1: Haiku, GPT-4o-mini (~$0.001)"]
  A -->|Write / implement / review / synthesize| T2["Tier 2: Sonnet, GPT-4o (~$0.01)"]
  A -->|Reason / architect / judge / decompose| T3["Tier 3: Opus, o1 (~$0.10)"]
  
  T1 --> Q1{Quality sufficient?}
  Q1 -->|Yes| Done1[Use cheap model]
  Q1 -->|No| T2
  
  T2 --> Q2{Quality sufficient?}
  Q2 -->|Yes| Done2[Use balanced model]
  Q2 -->|No| T3

Tier Assignment Table

Task TypeTierModelsCost/CallWhy This Tier
Classify input type1Haiku, GPT-4o-mini~$0.001Deterministic categorization
Validate schema/format1Haiku, GPT-4o-mini~$0.001Mechanical checking
Format output / template1Haiku, GPT-4o-mini~$0.001Structured transformation
Extract structured data1Haiku, GPT-4o-mini~$0.001Pattern matching
Summarize text1-2Haiku → Sonnet~$0.001-0.01Short summaries: Haiku; nuanced: Sonnet
Write content/docs2Sonnet, GPT-4o~$0.01Creative quality matters
Implement code2Sonnet, GPT-4o~$0.01Correctness + style
Review code/diffs2Sonnet, GPT-4o~$0.01Needs judgment, not just pattern matching
Research synthesis2Sonnet, GPT-4o~$0.01Multi-source reasoning
Decompose ambiguous problem3Opus, o1~$0.10Requires deep understanding
Design architecture3Opus, o1~$0.10Complex system reasoning
Judge output quality3Opus, o1~$0.10Meta-reasoning about quality
Plan multi-step strategy3Opus, o1~$0.10Long-horizon planning

Three Routing Strategies

Strategy 1: Static Tier Assignment (Start Here)

Assign model by task type at DAG design time. No runtime logic. Gets 60-70% of possible savings.

nodes:
  - id: classify
    model: claude-haiku-4-5     # Tier 1: $0.001
  - id: implement
    model: claude-sonnet-4-5    # Tier 2: $0.01  
  - id: evaluate
    model: claude-opus-4-5      # Tier 3: $0.10

Strategy 2: Cascading (Try Cheap First)

Try the cheap model; if quality is below threshold, escalate. Adds ~1s latency but saves 50-80% on nodes where cheap succeeds.

1. Execute with Tier 1 model
2. Quick quality check (also Tier 1 — costs ~$0.001)
3. If quality ≥ threshold → done
4. If quality < threshold → re-execute with Tier 2

Best for nodes where you're genuinely unsure which tier is needed.

Strategy 3: Adaptive (Learn from History)

Record success/failure per task type per model. Over time, the router learns:

  • "Classification nodes always succeed on Haiku" → stay cheap
  • "Code review nodes fail on Haiku 40% of the time" → upgrade to Sonnet
  • "Architecture nodes succeed on Sonnet 90% of the time" → don't need Opus

Gets 75-85% savings after ~100 executions of training data.


Provider Selection

Once model tier is chosen, select the provider:

Model ClassProvider OptionsSelection Criteria
Haiku-classAnthropic, AWS BedrockLatency, regional availability
Sonnet-classAnthropic, AWS Bedrock, GCP VertexCost, rate limits
Opus-classAnthropicOnly provider
GPT-4o-classOpenAI, Azure OpenAIRate limits, compliance
Open-sourceOllama (local), Together.ai, FireworksCost ($0), latency, GPU availability

Cost Impact Example

10-node DAG, "refactor a codebase":

StrategyMixCostSavings
All Opus10× $0.10$1.00—
All Sonnet10× $0.01$0.1090%
Static tiers4× Haiku + 4× Sonnet + 2× Opus$0.2476%
Cascading6× Haiku + 3× Sonnet + 1× Opus$0.1486%
Adaptive (trained)Dynamic~$0.0892%

Anti-Patterns

Always Use the Best Model

Wrong: Route everything to Opus/o1 "for quality." Reality: 60%+ of typical DAG nodes are classification, validation, or formatting — tasks where Haiku performs identically to Opus. You're burning money.

Always Use the Cheapest Model

Wrong: Route everything to Haiku "for cost." Reality: Complex reasoning, architecture design, and quality judgment genuinely need stronger models. Haiku will produce plausible-looking but subtly wrong output on hard tasks.

Ignoring Latency

Wrong: Only optimizing for cost, ignoring that Opus takes 5-10x longer than Haiku. Reality: In a 10-node DAG, model choice affects total execution time as much as cost. Route time-critical paths to faster models.

No Feedback Loop

Wrong: Setting model tiers once and never adjusting. Reality: As models improve (Haiku gets smarter every generation), tasks that needed Sonnet last month may work on Haiku today. Record outcomes and adapt.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.