cost-management

name	cost-management
description	Use when controlling AI spend, token budgets, model routing, or workflow efficiency before scaling usage

Cost Management

Announce at start: "Following the cost-management skill — budget before build."

Core Rule

Every AI workflow needs a budget and attribution. Cost you cannot segment or cap will eventually surprise you.

Process

1. Identify the Workflow

Define what is being measured:

• Workflow name (code_review, test_generation, incident_triage, etc.)
• Team or owner
• Environment (dev, ci, staging, prod)
• Expected frequency

2. Set a Budget Before Use

Choose a default cap before requests start flowing:

Budget scope	Example
Per task	$0.25 per PR review
Per user	$100/month per developer
Per team	$2,000/month for CI workflows
Per environment	Stricter caps in development than production

3. Right-Size the Model

Route by task complexity, not habit:

Task	Default tier
Formatting or classification	Cheapest capable model
Code review or test generation	Mid-tier reasoning model
High-risk architecture or security analysis	Highest tier only when justified

4. Reduce Token Waste

Before spending more, shrink the context:

• Send only relevant files or functions
• Use diffs instead of full before/after files when possible
• Cap output tokens to the needed response length
• Summarize long history instead of replaying all of it
• Fix retry loops before upgrading model tier

5. Instrument Spend

Every request should carry attribution and outcome data:

• Team
• Workflow
• Environment
• Model tier
• Token or cost usage
• Retry count

6. Review and Adjust

Use review cadence, not only invoices:

• Watch for high-cost workflows with weak outcomes
• Downgrade models if quality is unchanged
• Add alerts at budget thresholds
• Stop or redesign workflows that thrash

Red Flags — STOP

Signal	Action
Most expensive model is the default for all tasks	Add explicit routing rules
No request tags for team/workflow/env	Add attribution before scaling usage
Full repositories sent for tiny tasks	Trim context and re-measure
Cost review happens only after the invoice arrives	Add pre-flight caps and threshold alerts
Same workflow retries many times with flat quality	Fix prompts/process before spending more

Verification Checklist

• Workflow owner and budget were defined
• Model tier matches task complexity
• Context was trimmed to the minimum useful scope
• Requests are tagged with team, workflow, and environment
• Alerts or review cadence exist for ongoing spend

Related Skills

When	Invoke
Need better prompts to reduce retries	prompting
Designing specialized agents with bounded scope	refactoring or planning
Keeping spend policy reflected in docs	doc-maintenance

Deep Reference

For principles, rationale, anti-patterns, and examples:

• guides/cost-token-management/cost-token-management.md
• guides/ai-agent-evaluation-metrics/ai-agent-evaluation-metrics.md
• guides/context-management/context-management.md