Menu
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/Miosa-osa/canopy --skill latsThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Showing SKILL.md
Evaluate retrieval and generation quality in RAG pipelines. Separate scoring for retrieval (recall, precision, MRR) and generation (faithfulness, relevance, completeness). End-to-end pipeline assessment with bottleneck identification. Triggers on: "eval rag", "rag evaluation", "retrieval evaluation", "rag quality", "rag metrics"
Design binary pass/fail LLM-as-Judge evaluators. Structured prompt engineering for evaluation: criteria definition, rubric construction, few-shot calibration, and bias mitigation. Produces a ready-to-deploy judge prompt with scoring instructions. Triggers on: "judge prompt", "llm judge", "evaluator prompt", "scoring prompt", "grading rubric"
Self-learning system based on SICA, VIGIL, and Mem0 patterns. Auto-triggers after task completion. Captures patterns, consolidates memory, generates skills, recovers from errors.
Based on SOC occupation classification
| name | lats |
| description | Language Agent Tree Search - Monte Carlo planning - 92.7% on HumanEval |
| trigger | complex planning, code generation, decision-making under uncertainty |
| priority | 1 |
| dynamic | false |
| created | 2026-01-26 |
Monte Carlo Tree Search combined with LLM reasoning. Achieved 92.7% pass@1 on HumanEval (SOTA).
Use for:
while not solved and budget > 0:
1. SELECT: Pick best node using UCT formula
2. EXPAND: Generate N candidate actions
3. SIMULATE: Execute actions, get environment feedback
4. REFLECT: Self-evaluate trajectory quality
5. BACKPROPAGATE: Update scores up the tree
UCT(node) = exploitation + C * sqrt(ln(N) / n)
= avg_score + exploration_bonus
Where:
- C = exploration constant (typically 1.41)
- N = parent visit count
- n = node visit count
Generate top-5 candidate actions in parallel using the Task tool.
"Given this trajectory and outcome:
Trajectory: [actions taken]
Result: [success/failure + details]
Rate this approach 1-10 and explain:
1. What worked well?
2. What went wrong?
3. How could it be improved?"
def backpropagate(node, score):
while node:
node.visits += 1
node.total_score += score
node = node.parent
LATS is activated by @master-orchestrator when:
LATS is compute-intensive (5-10x more LLM calls). Reserve for:
Based on ICML 2024 research - arXiv:2310.04406