m2ai-portfolio Context Cost Line Item Analyzer
Breaks down an agent session's token spend by driver category, separating context-loaded-before-first-prompt from output-generated-during-session. Identifies the largest context contributors and ranks reduction opportunities by token savings.
Trigger
Use when the user says "context cost breakdown", "context line items", "why is context so expensive", "decompose my token spend", "what's eating my context window", "context vs output analysis", or wants to understand why context -- not output -- is the dominant cost in their agent workflow.
Phase 1: Classify Spend
Separate total token spend into two buckets:
Context bucket (loaded before first useful token of output):
- Startup overhead: global CLAUDE.md, project CLAUDE.md, MEMORY.md files
- Skill descriptions: all SKILL.md
descriptionfrontmatter fields - MCP tool schemas: all registered MCP server tool definitions
- Hook configurations: PreToolUse, PostToolUse, Stop, SessionStart hooks
- System prompt: base instructions, persona, rules injected by the harness
- Prior conversation turns retained in context (compacted or not)
Output bucket (generated during the session):
- Tool call parameters and results
- Assistant text responses
- Thinking/reasoning tokens (if extended thinking enabled)
Ask the user to provide: current session token usage breakdown (available from Claude Code /cost or session stats), their CLAUDE.md chain, and their MCP server list. If unavailable, estimate from environment scan.
Phase 2: Environment Scan
If the user cannot provide a usage breakdown, estimate from the local environment:
Scan checklist:
[ ] Count CLAUDE.md files in chain (global + project + .claude/)
[ ] List all skills/ and count description field character lengths
[ ] List MCP servers from settings.json / settings.local.json
[ ] List hooks from settings files
[ ] Check MEMORY.md size
Estimate tokens: ~4 chars per token (rough heuristic for English prose).
Phase 3: Driver Ranking Table
Produce a ranked table of context drivers:
Context Cost Line Items
========================
Driver Tokens % of Context
----------------------------- ------- ------------
MCP tool schemas (N servers) XXK XX%
CLAUDE.md chain (N files) XXK XX%
Skill descriptions (N skills) XXK XX%
System prompt XXK XX%
Prior turns (compacted) XXK XX%
Hooks configuration XXK XX%
MEMORY.md files XXK XX%
----------------------------- ------- ------------
Total context load XXK 100%
Output tokens this session: XXK
Context:Output ratio: X.X:1
Diagnosis: [CONTEXT-HEAVY / BALANCED / OUTPUT-HEAVY]
A ratio above 3:1 (context to output) signals that context cost is the problem, not output volume.
Phase 4: Reduction Recommendations
For each top driver, propose a targeted intervention:
| Driver | Reduction Action | Estimated Savings |
|---|---|---|
| MCP schemas | Disable unused MCP servers | Up to X tokens/session |
| Skill descriptions | Shorten verbose descriptions | Up to X tokens/session |
| CLAUDE.md chain | Consolidate or trim rules | Up to X tokens/session |
| Prior turns | Reduce compaction threshold | Up to X tokens/session |
| System prompt | Defer boilerplate to first user turn | Up to X tokens/session |
Rank by savings potential. Do not recommend disabling something that is actively referenced by a running workflow.
Phase 5: Report Output
Context Cost Analysis — [DATE]
==============================
[Driver ranking table from Phase 3]
Top 3 Reduction Opportunities:
1. [Action] — saves ~XXK tokens/session (~$X.XX/month at current usage)
2. [Action] — saves ~XXK tokens/session
3. [Action] — saves ~XXK tokens/session
Context:Output ratio: X.X:1 ([status])
If the ratio is already under 2:1, report that context is not the problem and redirect attention to output efficiency.
Integration Notes
- Pairs with
token-burn-auditor(which measures total overhead) andboot-tax-monitor(which tracks startup cost specifically) - Use after a heavy session to understand where the spend went before optimizing
- Reduction actions can be handed to
compensating-complexity-auditorto check whether any MCP or skill flagged for removal is actively load-bearing - Can be run on a per-project or per-agent basis by scoping the CLAUDE.md chain and settings files
Source Attribution
Technique derived from Nate's Newsletter (2026-05-10): "Executive Briefing: Six announcements in 48 hours just changed how enterprise AI gets bought" -- Nate's observation that "context, not tokens, is the line item ruining agent economics" and that capping usage kills the use case without fixing the cause.