New /context command is useful

[stephenfeather]

Been too busy fighting with Anthropic over mismatches between usage reports from claude-code vs web/desktop to catch the announcement regarding the new /context command.

(I'm out of usage on a Pro account until Friday at 3 pm my time. Learned you can not 'buy' your way out of the 5-hour limits, but not the weekly. :()

Anyway, here is a /context response from an active dev project. I should note that this was the first prompt, from the first run of Claude, in a fresh iTerm terminal.

The Takeaways

Looking through this, here are some things I gleaned.

1. My rule file for preferred cli tools must be a bit too wordy. At 1500 tokens, modern-cli-tools.md stands out against the couple of hundred tokens the other rules use.

2. My skill wordpress-php-dev must have an excessively large header that needs to be looked at. For a skill that won't ever be used in this particular project, it's nearly 2.5x larger than its next neighbor and 4x larger than the one after that.

3. Despite having tldr-mcp setup in settings.json and in mcp_config.json, it doesn't appear to be loading. (It did load when defined in claude.json?) Morph is not being loaded from mcp_config.json with a disabled:false, its coming from claude.json.

4. With item 3 in mind, it would be nice not to load tldr-cli from rules if the mcp server isn't up. (or any other similar scenerio)

The Data

❯ /context
  ⎿   Context Usage
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   claude-opus-4-5-20251101 · 32k/200k tokens (16%)
     ⛀ ⛀ ⛁ ⛁ ⛁ ⛀ ⛁ ⛀ ⛶ ⛶   ⛁ System prompt: 2.8k tokens (1.4%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ System tools: 17.7k tokens (8.9%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ MCP tools: 1.1k tokens (0.5%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Custom agents: 623 tokens (0.3%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Memory files: 7.3k tokens (3.6%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Skills: 2.1k tokens (1.1%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ ⛝ ⛝   ⛁ Messages: 65 tokens (0.0%)
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝   ⛶ Free space: 123k (61.7%)
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝   ⛝ Autocompact buffer: 45.0k tokens (22.5%)

     MCP tools · /mcp
     └ mcp__morph-mcp__edit_file: 636 tokens
     └ mcp__morph-mcp__warpgrep_codebase_search: 421 tokens

     Custom agents · /agents

     User
     └ agentica-agent: 36 tokens
     └ review-agent: 33 tokens
     └ plan-agent: 24 tokens
     └ plan-reviewer: 24 tokens
     └ debug-agent: 24 tokens
     └ braintrust-analyst: 23 tokens
     └ memory-extractor: 23 tokens
     └ onboard: 23 tokens
     └ validate-agent: 23 tokens
     └ scribe: 22 tokens
     └ context-query-agent: 21 tokens
     └ oracle: 21 tokens
     └ session-analyst: 21 tokens
     └ research-codebase: 20 tokens
     └ chronicler: 19 tokens
     └ kraken: 19 tokens
     └ profiler: 18 tokens
     └ architect: 17 tokens
     └ herald: 17 tokens
     └ atlas: 17 tokens
     └ judge: 16 tokens
     └ arbiter: 16 tokens
     └ maestro: 16 tokens
     └ phoenix: 16 tokens
     └ sleuth: 16 tokens
     └ scout: 16 tokens
     └ spark: 15 tokens
     └ pathfinder: 15 tokens
     └ aegis: 14 tokens
     └ critic: 13 tokens
     └ surveyor: 13 tokens
     └ liaison: 12 tokens

     Memory files · /memory
     └ ~/.dotfiles/claude/rules/claim-verification.md: 508 tokens
     └ ~/.dotfiles/claude/rules/no-haiku.md: 73 tokens
     └ ~/.dotfiles/claude/rules/hook-auto-execute.md: 153 tokens
     └ ~/.dotfiles/claude/rules/dynamic-recall.md: 1.3k tokens
     └ ~/.dotfiles/claude/rules/destructive-commands.md: 490 tokens
     └ ~/.dotfiles/claude/rules/agent-model-selection.md: 228 tokens
     └ ~/.dotfiles/claude/rules/tldr-cli.md: 1.2k tokens
     └ ~/.dotfiles/claude/rules/use-scout-not-explore.md: 226 tokens
     └ ~/.dotfiles/claude/rules/proactive-delegation.md: 624 tokens
     └ ~/.dotfiles/claude/rules/proactive-memory-disclosure.md: 343 tokens
     └ ~/.dotfiles/claude/rules/agent-memory-recall.md: 175 tokens
     └ ~/.dotfiles/claude/rules/modern-cli-tools.md: 1.5k tokens
     └ ~/.dotfiles/claude/rules/cross-terminal-db.md: 382 tokens

     Skills · /skills

     User
     └ wordpress-php-dev: 138 tokens
     └ compound-learnings: 55 tokens
     └ fix: 34 tokens
     └ discovery-interview: 33 tokens
     └ debug-hooks: 32 tokens
     └ repo-research-analyst: 31 tokens
     └ tdd-migration-pipeline: 30 tokens
     └ shapely-compute: 29 tokens
     └ agentica-sdk: 28 tokens
     └ tldr-code: 28 tokens
     └ hook-developer: 27 tokens
     └ math: 27 tokens
     └ braintrust-tracing: 27 tokens
     └ workflow-router: 26 tokens
     └ test-driven-development: 26 tokens
     └ research-agent: 26 tokens
     └ slash-commands: 26 tokens
     └ tdd: 26 tokens
     └ math-router: 25 tokens
     └ pint-compute: 25 tokens
     └ agentica-prompts: 25 tokens
     └ plan-agent: 25 tokens
     └ SonarQube: 25 tokens
     └ help: 25 tokens
     └ implement_task: 24 tokens
     └ validate-agent: 24 tokens
     └ explore: 24 tokens
     └ agentica-server: 24 tokens
     └ research-external: 23 tokens
     └ sub-agents: 23 tokens
     └ release: 23 tokens
     └ mcp-chaining: 23 tokens
     └ implement_plan_micro: 23 tokens
     └ continuity_ledger: 22 tokens
     └ tdd-migrate: 22 tokens
     └ skill-upgrader: 22 tokens
     └ agentica-claude-proxy: 21 tokens
     └ math-help: 21 tokens
     └ resume_handoff: 21 tokens
     └ morph-apply: 21 tokens
     └ implement_plan: 21 tokens
     └ agentica-infrastructure: 21 tokens
     └ gemini-image-generator: 21 tokens
     └ review: 21 tokens
     └ cli-reference: 21 tokens
     └ refactor: 21 tokens
     └ research: 20 tokens
     └ tldr-stats: 20 tokens
     └ perplexity-search: 20 tokens
     └ describe_pr: 20 tokens
     └ test: 20 tokens
     └ create_handoff: 20 tokens
     └ premortem: 20 tokens
     └ build: 20 tokens
     └ prove: 19 tokens
     └ recall-reasoning: 19 tokens
     └ migrate: 19 tokens
     └ commit: 18 tokens
     └ repoprompt: 18 tokens
     └ firecrawl-scrape: 18 tokens
     └ skill-developer: 18 tokens
     └ onboard: 18 tokens
     └ github-search: 18 tokens
     └ debug: 18 tokens
     └ qlty-check: 17 tokens
     └ security: 17 tokens
     └ mot: 17 tokens
     └ ast-grep-find: 17 tokens
     └ morph-search: 17 tokens
     └ braintrust-analyze: 16 tokens
     └ graceful-degradation: 16 tokens
     └ search-router: 16 tokens
     └ reference-sdk: 16 tokens
     └ completion-check: 16 tokens
     └ nia-docs: 16 tokens
     └ dead-code: 15 tokens
     └ parallel-agent-contracts: 14 tokens
     └ explicit-identity: 13 tokens
     └ no-polling-agents: 13 tokens
     └ router-first-architecture: 13 tokens
     └ agentica-spawn: 13 tokens
     └ opc-architecture: 12 tokens
     └ complete-skill: 12 tokens
     └ agent-context-isolation: 12 tokens
     └ qlty-during-development: 12 tokens
     └ idempotent-redundancy: 11 tokens
     └ background-agent-pings: 11 tokens
     └ observe-before-editing: 11 tokens
     └ agent-orchestration: 11 tokens
     └ parallel-agents: 11 tokens
     └ index-at-creation: 10 tokens
     └ async-repl-protocol: 10 tokens
     └ skill-development: 10 tokens
     └ modular-code: 10 tokens
     └ agentic-workflow: 10 tokens
     └ llm-tuning-patterns: 10 tokens
     └ search-hierarchy: 10 tokens
     └ environment-triage: 9 tokens
     └ no-task-output: 9 tokens
     └ search-tools: 9 tokens
     └ wiring: 7 tokens
     └ mcp-scripts: 7 tokens
     └ git-commits: 7 tokens
     └ hooks: 7 tokens

[stephenfeather]

It would only be fair to compare the CCv3 load with a clean Claude setup.
19k clean vs 32k CC3.

But look at what the additional 13k gets you, am I right?

/context
  ⎿   Context Usage
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛀   claude-opus-4-5-20251101 · 19k/200k tokens (10%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ System prompt: 2.7k tokens (1.4%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ System tools: 16.7k tokens (8.4%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Messages: 8 tokens (0.0%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛶ Free space: 136k (67.8%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛝ Autocompact buffer: 45.0k tokens (22.5%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛝ ⛝ ⛝
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝
     ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝ ⛝

[parcadei]

yeah this needs to go down hahaha

note to self to carefully organise experimental agents vs actual prod agents properly in the .gitignore

will focus on bringing this down to ~20-22k like v2

good spot!