arbor

Autonomously improve a real artifact (code, training recipe, agent harness, data pipeline, prompt) against an objective and an evaluator, using Hypothesis Tree Refinement (HTR) from the Arbor paper. Use this whenever someone wants to iteratively optimize something over many experiments without overfitting — e.g. "get my model's eval score up", "improve this agent/harness", "tune this pipeline", "beat the baseline on this benchmark", "run a search over approaches and keep the best", "do an MLE-bench / Kaggle-style optimization", or any long-horizon "make this artifact better and don't just memorize the dev set" task. Trigger it even when the user doesn't say "Arbor" or "hypothesis tree" but describes repeated experiment-and-evaluate loops, branching exploration of competing ideas, or worries about a dev/test gap. Runs Claude itself as the coordinator with subagent executors in isolated git worktrees; for the standalone `arbor` CLI tool see references/arbor-upstream.md.

Quality

92%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Quality

Content

85%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is a strong, actionable orchestration guide with a clearly sequenced six-step loop, a validation merge gate, and well-organized one-level-deep references to real bundle files. Its one weak spot is conciseness — rhetorical flourishes and embedded paper statistics add tokens without adding operational value.

Suggestions

Trim rhetorical padding such as 'lessons accumulate instead of evaporating' and 'not a victory lap' to plain operational statements.

Move the MLE-Bench Lite ablation numbers (54.5% / 63.6% / 81.8%) into references/htr-methodology.md and keep only the takeaway ('insight propagation drives most of HTR's gains') in the body.

Consider condensing the Principles section, since several points restate guidance already given in the step-by-step loop.

Dimension	Reasoning	Score
Conciseness	Mostly efficient and dense with operational content, but includes rhetorical padding ('lessons accumulate instead of evaporating', 'not a victory lap') and inessential paper ablation numbers ('54.5% vs. 63.6% … against 81.8%') that could be tightened, matching the 'mostly efficient but includes some unnecessary explanation' anchor rather than the lean top anchor.	2 / 3
Actionability	Provides fully executable, copy-paste-ready commands throughout — `tree.py init`, `add-node`, `set-evidence`, `propagate`, `prune`, `merge` — each with concrete arguments and example values, matching the 'fully executable code/commands' anchor.	3 / 3
Workflow Clarity	The six-step coordinator loop (Observe → Ideate → Select → Dispatch → Backpropagate → Decide) is explicitly sequenced with a held-out merge-gate validation checkpoint and feedback on gate rejection, matching the 'clear sequence with explicit validation steps' anchor.	3 / 3
Progressive Disclosure	The body is an operational overview pointing to four real one-level-deep reference files (htr-methodology.md, executor-brief.md, report-template.md, arbor-upstream.md), each clearly signaled inline and listed in a Reference files section, matching the 'clear overview with well-signaled one-level-deep references' anchor.	3 / 3
	Total	11 / 12 Passed

Description

100%

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is comprehensive, concrete, and rich with natural trigger phrasings, with an explicit 'Use when' clause and a well-scoped niche. Its only weakness is verbosity: the dense single paragraph could be tightened without losing substance.

Dimension	Reasoning	Score
Specificity	Names multiple concrete actions and artifact types — 'improve a real artifact (code, training recipe, agent harness, data pipeline, prompt)', 'iteratively optimize', 'run a search over approaches and keep the best' — matching the 'lists multiple specific concrete actions' anchor.	3 / 3
Completeness	Explicitly answers both what (autonomous optimization loop using HTR) and when via an explicit 'Use this whenever someone wants to iteratively optimize…' clause, matching the top anchor.	3 / 3
Trigger Term Quality	Rich coverage of natural user phrasings such as 'get my model's eval score up', 'tune this pipeline', 'beat the baseline on this benchmark', and 'do an MLE-bench / Kaggle-style optimization', matching the 'good coverage of natural terms' anchor.	3 / 3
Distinctiveness Conflict Risk	Occupies a clear niche — long-horizon iterative artifact optimization under a dev/test evaluator split — with distinct triggers unlikely to fire for unrelated skills.	3 / 3
	Total	12 / 12 Passed

Validation

100%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository: K-Dense-AI/scientific-agent-skills
Path: skills/arbor/SKILL.md
Commit: 757b63b

Reviewed: about 11 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.