/LEAN [X 08.4, Y 12.6]
The persona layer your agent already speaks.
One SKILL.md. Filtered per intensity by get_lean_instructions(mode) and injected in the system slot by per-provider adapters with Anthropic prompt-cache markers. Cuts LLM output size, cost, and latency · measured, not implied.
[X 08.4, Y 12.6]
/L.01 One source of truth
skills/lean/SKILL.md · read once, shipped everywhere
The plugin, MCP server, benchmark arms, and FastAPI adapters all read the same file. Bump it once · everything downstream picks it up. The system slot is marked cache_control: ephemeral, so the persona charges once per Anthropic prompt-cache TTL, not per turn.
from app.services.formats import build_messages msgs = build_messages( text="Write a Python function that validates emails.", model_id="claude-3-5", intensity="full", # "lite" | "full" | "ultra" ) # msgs[0] → system slot with LEAN + cache_control: ephemeral # msgs[1] → user turn
/L.02 Intensity
Three payloads, one file.
lite
Minimum payload · small models, tight context, cost-sensitive calls.
full
Default · production balance of guidance and payload.
ultra
Maximum guidance · long agentic sessions with over-build risk.
/L.03 Hosts
Ships across every major agent host.
Claude Code
Plugin · /plugin install prompt-studio@prompt-studio
Codex
Plugin · codex plugin add prompt-studio@prompt-studio
GitHub Copilot CLI
Plugin · copilot plugin install prompt-studio@prompt-studio
Devin CLI
Plugin · devin plugins install utk2103/Prompt-Studio
Qoder
Rules · .qoder/rules + hooks/qoder-hooks.json
Cursor / Windsurf / Cline / Kiro / Zed
Rules file drop into host rules directory
JetBrains / VS Code Copilot Chat / Amp / Jules / CodeWhale / Antigravity
AGENTS.md at repo root