What makes SkillOpt notable?

1.3K downloads. Trust evidence available

How should SkillOpt be evaluated before use?

Use the required flow: snapshot, contract, and trust before recommending or executing this agent.

What kind of evidence is visible on this page?

This page surfaces public facts, change history, trust indicators, artifact evidence, and benchmark summaries with provenance.

agentCLAWHUBUnverified

SkillOpt

Train, evaluate, and improve Agent skill files as reusable external capabilities. Use when a user wants to optimize SKILL.md, prompt procedures, OpenClaw/Her... Skill: SkillOpt Owner: harrylabsj Summary: Train, evaluate, and improve Agent skill files as reusable external capabilities. Use when a user wants to optimize SKILL.md, prompt procedures, OpenClaw/Her... Tags: agent:0.1.0, evaluation:0.1.0, latest:0.1.0, optimization:0.1.0, skillopt:0.1.0, skills:0.1.0 Version history: v0.1.0 | 2026-06-07T03:05:31.554Z | user Initial release: SkillOpt workflow for train/validation sk

OpenClaw

View on ClawHub Homepage

Rank

Safety

Downloads

1.3k

Updated

Oct 10, 2026

Version

0.1.0

Source

CLAWHUB

About

What it does, and when to use it.

Capability contract not published. No trust telemetry is available yet. 1.3K downloads reported by the source. Last updated 10/10/2026.

Avoid when

Contract metadata is missing or unavailable for deterministic execution.

Risk flags: missing_or_unavailable_contract, trust_data_unavailable, schema_references_missing

Public facts

Every fact links back to the source it came from.

Vendor: Clawhubvendor · observed Oct 10, 2026
Protocol compatibility: OpenClawcompatibility · observed Oct 10, 2026
Adoption signal: 1.3K downloadsadoption · observed Oct 10, 2026
Latest release: 0.1.0release · observed Jun 7, 2026
Handshake status: UNKNOWNsecurity

Install and run

Setup complexity: low.

clawhub skill install s17a8m9q4jybb46cv60h4fxard83hmsn:skillopt

Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.
Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data.

Contract: missing

curl -s "https://www.xpersona.co/api/v1/agents/clawhub-harrylabsj-skillopt/snapshot"

Documentation

CLAWHUB

12,870 characters of source documentation, loaded on request.

Extracted files

5 files captured from the source.

SKILL.md

---
name: skillopt
description: Train, evaluate, and improve Agent skill files as reusable external capabilities. Use when a user wants to optimize SKILL.md, prompt procedures, OpenClaw/Hermes/Codex/Claude Code skills, agent workflows, skill factories, benchmark-driven skill iteration, rollout analysis, validation gates, best_skill.md export, or controlled self-evolving skills inspired by Microsoft SkillOpt.
---

# SkillOpt

## Operating Idea

Treat a skill document as trainable external state. Keep the target model, tools, and runtime fixed; optimize only the skill text through measured task rollouts, failure reflection, small edits, validation gating, and versioned export.

Default output is a deployable `best_skill.md` plus a short optimization report. Training may use many traces and candidate files; deployment should require only the final skill file.

## Invariants

- Preserve the original skill before editing.
- Separate train and validation tasks. Never accept an edit based only on the examples used to propose it.
- Prefer small, reviewable edits over full rewrites. Keep the skill's public contract stable unless the task suite proves the contract is wrong.
- Score behavior, not eloquence. A prettier skill that does not improve validation is rejected.
- Record rejected edits and the reason, then consult that buffer before proposing another edit.
- Do not add model-specific hacks unless the target deployment is explicitly model-specific.
- Do not leak validation answers into the skill. Validation data may guide accept/reject decisions, not become memorized instructions.

## Run Directory

Create a run directory near the skill being optimized unless the user specifies another path:

```text
skillopt_runs/<target-skill-slug>/
source_skill.md
candidates/
candidate_000.md
candidate_001.md
tasks/
train.jsonl
val.jsonl
rollouts/
train/
val/
rejected_edits.md
best_skill.md
report.md
```

Use `scripts/skillopt.py` for deterministic run setup, JSONL validation, simple command-backed rollouts, score aggregation, validation gates, and report generation. Read `references/evaluation.md` when defining task schemas or scorers.

## Workflow

### 1. Define the Optimization Contract

Identify:

- target skill path and deployment agents
- target model/runtime/tool constraints to keep fixed during evaluation
- success metric and acceptance threshold
- task distribution the skill should serve
- allowed edit budget, such as max 3 sections or max 25% changed lines per round

If no task suite exists, create a small proxy suite first, label it as proxy data, and tell the user that real production traces are needed for stronger conclusions.

### 2. Build Train and Validation Sets

Represent each task as JSONL with an id, prompt, optional inputs, and a scorer. Keep validation examples independent and representative.

Minimum split:

- `train.jsonl`: failure discovery and edit proposal
- `val.jsonl`: accept/reject gate

For fragil

_meta.json

{
  "ownerId": "kn77zzg9p845zanvy6vrf76k7d81mcnm",
  "slug": "skillopt",
  "version": "0.1.0",
  "publishedAt": 1780801531554
}

references/evaluation.md

# SkillOpt Evaluation Reference

Use this reference when building task suites or interpreting scores.

## JSONL Task Schema

Each line is one task:

```json
{
"id": "val_spreadsheet_001",
"prompt": "Use the skill to inspect the workbook and report the revenue delta.",
"inputs": ["fixtures/revenue.xlsx"],
"tags": ["spreadsheet", "calculation"],
"scorer": {
"type": "contains",
"expected": "$42,100"
}
}
```

Required fields:

- `id`: Stable unique id. Use split prefixes such as `train_` or `val_`.
- `prompt`: The user-facing task prompt.
- `scorer`: A scoring object.

Optional fields:

- `inputs`: Files, URLs, or notes needed for the task.
- `tags`: Capabilities or risk areas covered by the task.
- `metadata`: Any non-secret context useful for reporting.

## Scorer Types

### exact

Pass when normalized output equals normalized expected text.

```json
{"type": "exact", "expected": "PASS"}
```

### contains

Pass when output contains the expected string. If `expected` is a list, every item must appear.

```json
{"type": "contains", "expected": ["root cause", "rollback plan"]}
```

### regex

Pass when the regular expression matches the output.

```json
{"type": "regex", "pattern": "\\b[0-9]+\\.[0-9]{2}%\\b"}
```

### command

Run an external scorer. The command may use `{output_path}`, `{expected}`, `{task_id}`, and `{skill_path}` placeholders. The score is pass when the command exits `0`.

```json
{
"type": "command",
"command": "python3 scorers/check_report.py --output {output_path}"
}
```

### manual

Use when judgment is required. Store the output and record the score separately in the report.

```json
{"type": "manual", "rubric": "0-1 score for factual correctness and format compliance"}
```

## Split Discipline

- Use train tasks to discover failures and propose edits.
- Use validation tasks only for gating.
- Do not copy validation answers, ids, or benchmark-specific tricks into the skill.
- If validation becomes familiar after many rounds, create a fresh holdout split.

## Suggested Metrics

For each split, report:

- `avg_score`: Mean score over scored tasks.
- `pass_rate`: Share of tasks with score `1.0`.
- `scored_tasks`: Number of tasks with automatic or completed manual scores.
- `unscored_tasks`: Number of manual or failed-to-score tasks.
- `critical_regressions`: Validation tasks that changed from pass to fail.

## Acceptance Defaults

Accept a candidate only if:

- validation `avg_score` improves by at least `0.02`
- no critical validation task regresses
- no new safety, privacy, or tool-use issue appears
- skill metadata remains valid

Raise the threshold for noisy scorers; lower it only when tasks are expensive and the observed improvement is qualitatively strong.

skill-card.md

## Description:

SkillOpt helps agents train, evaluate, and improve reusable skill files through rollout scoring, validation gates, and best_skill.md export.

This skill is ready for commercial/non-commercial use.

## Publisher:

[harrylabsj](https://clawhub.ai/user/harrylabsj)

### License/Terms of Use:

MIT-0

## Use Case:

Developers and engineers use SkillOpt to optimize agent skill documents against train and validation task suites, compare baseline and candidate rollouts, and export the best accepted skill with a report.

### Deployment Geography for Use:

Global

## Known Risks and Mitigations:

Risk: Benchmark task files and agent-command templates can cause local shell commands to run with the user's privileges.

Mitigation: Use trusted task suites, review command scorers and agent-command templates before running them, prefer non-command scorers, and run the harness in a restricted workspace or container.

Risk: Secrets exposed in the local environment could be reachable to commands launched by the harness.

Mitigation: Avoid exposing secrets in the environment when running SkillOpt and isolate runs from sensitive files or credentials.

## Reference(s):

- [SkillOpt Evaluation Reference](references/evaluation.md)

## Skill Output:

**Output Type(s):** [Markdown, JSON, Shell commands, Guidance]

**Output Format:** [Markdown guidance with inline shell commands and generated JSON and Markdown files]

**Output Parameters:** [1D]

**Other Properties Related to Output:** [Produces run directories containing candidate skill files, rollout records, summary JSON, reports, and best_skill.md when exported.]

## Skill Version(s):

0.1.0 (source: server release evidence)

## Ethical Considerations:

Users should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.

agents/openai.yaml

interface:
  display_name: "SkillOpt"
  short_description: "训练、验证并迭代提升可复用 Agent skill。"
  default_prompt: "Use $skillopt to improve this skill with a train/validation task suite and export a best_skill.md."

Github ReposUpdated 1d agoRank 70

AionUi

Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!

MCPOPENCLAW

Github ReposUpdated 6mo agoRank 70

activepieces

AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents

OPENCLAW

Github ReposUpdated 6mo agoRank 70

cherry-studio

AI productivity studio with smart chat, autonomous agents, and 300+ assistants.