← ClaudeAtlas

agent-kitlisted

Agent toolkit, one skill with sub-skills for running Claude Code agents cheaply. (1) cost: route each agent to the cheapest model tier that can finish it, give it a checkable finish line, skip agents the task doesn't need, with a PreToolUse hook that nudges Agent calls that set no model; (2) prompt-cache: keep agent prompt prefixes byte-identical so the prompt cache hits, with a uv-run script that builds a fingerprinted project outline, writes agent briefs and reports per-agent hit rate. Use WHENEVER you are about to spawn subagents, forks, parallel agents or a Workflow, run /loop or ScheduleWakeup, pick a model for an agent, or the user says "save tokens", "agent cost", "which model", "prompt caching", "cache hit", "token usage", "ลด token". Read the sub-skill's SKILL.md before acting.
th3r-dev/th3-skills · ★ 2 · AI & Automation · score 64
Install: claude install-skill th3r-dev/th3-skills
# agent-kit: sub-skills | When the work is… | Read | Scripts | |---|---|---| | spawning agents, "which model for this agent", "too many agents", "agent cost" | [`cost/SKILL.md`](cost/SKILL.md) | `cost/scripts/agent_guard.py` (PreToolUse hook, `--selftest`) | | fanning out agents that share context, /loop, "cache hit rate", "prompt caching" | [`prompt-cache/SKILL.md`](prompt-cache/SKILL.md) | `prompt-cache/scripts/pcache.py` (`build` · `brief` · `stats` · `--self-test`) | They chain: `cost` decides whether to spawn and on which tier; `prompt-cache` makes the briefs those agents get share one cached prefix, then measures it. ## Rules every sub-skill keeps - **No agent is free.** Count the brief, the re-read and the summary against doing it inline. - **Measure, don't estimate.** Report hit rates and token counts from the session logs (`pcache.py stats`). - **Never block the user's work.** The hook nudges; it doesn't refuse a call. ## Tests ```bash cd skills/agent-kit uv run --script cost/scripts/agent_guard.py --selftest uv run --script prompt-cache/scripts/pcache.py --self-test ```