te

skills(claude-token-efficiency) (te plugin)

Keep Claude Code's context window lean on long sessions and large codebases.

The core idea: the supervisor agent's context is RAM, not disk. Anything long, noisy, or iterative should happen outside the supervisor's window — in a subagent, in a scratch file, or trimmed before it lands.

Skills

Skill	What it does
token-efficient-agent	Meta/orchestrator — activates all three patterns at task start
delegate-work	Spawn subagents for search, iteration loops, and verification. Enforces structured return contracts.
tier-model	Decision table: Haiku for mechanical work, Sonnet for judgment, Opus only when Sonnet stalls
scratch-context	Externalize notes, decisions, and dead ends to .agent-scratch/ so they survive across turns

Install

Option A — Plugin (recommended)

Install the plugin so skills are available everywhere and the SessionStart hook auto-injects scratch notes:

/plugin marketplace add MuktadirHassan/skills
/plugin install te@skills

Skills are then available as /te:token-efficient-agent, /te:delegate-work, etc.

Then per project, run two setup commands:

# 1. Add operating rules so Claude applies the patterns automatically
curl -fsSL https://raw.githubusercontent.com/MuktadirHassan/skills/main/templates/CLAUDE.md -o CLAUDE.md

# 2. Enable scratch note injection
mkdir .agent-scratch && echo ".agent-scratch/" >> .gitignore

Without step 1, the skills exist but Claude won't apply them unless you invoke them manually.

Option B — Standalone skills (no hooks)

Copy the skills into your global ~/.claude/skills/:

git clone https://github.com/MuktadirHassan/skills /tmp/te
cp -r /tmp/te/skills/* ~/.claude/skills/

Restart Claude Code. Skills show up in /help as /token-efficient-agent, /delegate-work, etc.

Then add operating rules to any project you want them active in:

curl -fsSL https://raw.githubusercontent.com/MuktadirHassan/skills/main/templates/CLAUDE.md -o CLAUDE.md

Note: with Option B, the scratch-context skill works but you have to read .agent-scratch/ manually — there's no auto-injection hook.

Templates

File	Purpose
templates/CLAUDE.md	Drop-in operating rules (~5 lines). Copy to your project or ~/.claude/CLAUDE.md
templates/SYSTEM_PROMPT.md	Extended rules: model tiering table, failure modes. Paste the sections you want.
templates/return-contract.md	Copy-paste subagent reply schemas (default, iterator, scout)

Recommended companions

• CodeGraph — pre-indexed knowledge graph for structural code queries ("who calls X", "what does Y affect"). One MCP call instead of 30 greps. Ships its own Claude Code integration.
• RTK — trims verbose Bash output (60–90% savings on git, npm, etc.) before the model sees it. Ships its own hook.

Repo layout

.claude-plugin/plugin.json        plugin manifest (name: te)
hooks/
  hooks.json                      SessionStart registration
  load-scratch.py                 injects .agent-scratch/ contents on session start
skills/
  token-efficient-agent/SKILL.md  meta/orchestrator
  delegate-work/SKILL.md          subagent patterns and return contract
  tier-model/SKILL.md             model selection decision table
  scratch-context/SKILL.md        externalize state to .agent-scratch/
templates/
  CLAUDE.md                       drop-in operating rules
  SYSTEM_PROMPT.md                extended rules
  return-contract.md              subagent reply schemas

License

MIT. See LICENSE.