Guide · Agent skill
How to use the Meta-Optimize Skill Tuner agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Meta-Optimize is an agent skill from the ARIS research harness that reads logs of how the harness has been used and proposes changes to its own skill prompts, reviewer instructions, default parameters and workflow ordering. It is for developers who maintain or customize ARIS and have accumulated enough usage history to make data-driven tweaks. By the end of this guide you will know what the skill does, how to install it, how to trigger it, and what it needs before it produces useful output.
Meta-Optimize Skill Tunerwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Install the ZIP. Download the skill ZIP from this page and unzip it into ~/.claude/skills/ for personal use or .claude/skills/ inside a project. Alternatively, upload it under Settings → Capabilities → Skills in the Claude apps.
- 2Enable the logging hook. Copy templates/claude-hooks/meta_logging.json into your project's .claude/settings.json or merge its hooks section.
- 3Build up usage history. Run ARIS workflows normally until the event log has at least 5 logged skill invocations in .aris/meta/events.jsonl.
- 4Run the analysis. Invoke the skill with /meta-optimize, optionally followed by a target skill name, or ask in natural language such as "meta optimize".
- 5Read the report. Open the report written under .aris/meta/ and check the proposed changes against your own experience of the workflows.
- 6Land approved patches. Run /meta-apply yourself for any staged patches you approve. This human step is the landing gate.
What the Meta-Optimize skill makes the agent do
Meta-Optimize analyzes accumulated ARIS usage logs and proposes optimizations to the harness itself, not to the research artifacts the harness produces. The harness components it targets include SKILL.md prompts, default parameters such as difficulty or round limits, convergence rules for the review loop, and the order of skills within a workflow.
It only proposes changes. The skill writes its report and staged patches under .aris/meta/ and does not edit skill files. A separate, human-invoked skill, /meta-apply, must land any approved patches.
- Optimizable: SKILL.md prompts, default parameters, convergence rules, workflow ordering.
- Cautious: artifact schemas such as EXPERIMENT_LOG.md.
- Not optimized: MCP bridge config (infrastructure) and the research papers, code and experiments themselves.
Install the skill from ZIP or the Claude apps
Download the ZIP from this page and unzip it into one of two locations. For a personal install, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.
If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.
- Personal: unzip into ~/.claude/skills/
- Project only: unzip into .claude/skills/ in your project root
- Claude apps: Settings → Capabilities → Skills, then upload
Enable the meta_logging hook before you run it
The skill needs an event log to analyze. Copy templates/claude-hooks/meta_logging.json from the ARIS repository into your project's .claude/settings.json, or merge its hooks section into the file if one already exists. Without this hook there is no log, and the skill stops with an error.
The skill reads .aris/meta/events.jsonl. It needs at least 5 logged skill invocations before the analysis is useful. With fewer, it reports insufficient data and asks you to keep using ARIS and rerun later.
- Hook file: templates/claude-hooks/meta_logging.json
- Settings file: .claude/settings.json
- Event log: .aris/meta/events.jsonl
- Minimum: 5 skill invocations
Trigger the skill with natural requests
You can invoke the skill with a slash command and an optional target, or by asking in natural language. The skill's argument hint is a target skill name or all. The description lists trigger phrases including "meta optimize", "improve skills", "优化技能" and "分析使用记录".
Ask for a specific target when you only want to improve one skill, or ask for all to analyze the whole harness.
- "meta optimize" or "improve skills" to run a general analysis
- "分析使用记录" to analyze usage records
- "optimize the auto-review-loop defaults" style requests scope the run to one component
/meta-optimize [target-skill-or-all]What the analysis looks at
The analysis covers usage frequency, failures, convergence and human intervention. Frequency looks at which skills and slash commands are used most and which parameter overrides recur, since frequent overrides can suggest bad defaults.
Failure analysis counts which tools fail and which error patterns repeat, such as out-of-memory, import, compilation or timeout errors. Convergence analysis, for the auto-review-loop skill, looks at average rounds to reach threshold, the shape of score trajectories and whether users override difficulty mid-run.
- Frequency: most-invoked skills, common slash commands, common parameter overrides
- Failures: failing tools, repeating error patterns, auto-debug retries per run
- Convergence: rounds to threshold, score trajectory shape, mid-run difficulty overrides
- Interventions: where users interrupt a workflow
Safety notes: scripts, network and write access
This skill runs scripts and uses the network. Its tool list includes Bash and several optional reviewer backends through MCP, such as Codex, Antigravity, Grok or a Claude review MCP for cross-model review. Those reviewers are optional; the skill's reviewer routing falls back to the Claude Code default when no selection is set.
The skill is designed as a read-only producer. It has no Write or Edit tool, and its legitimate Bash writes go only to .aris/meta/. The source is explicit that Bash is filtered, not sandboxed: an optional corpus_write_guard PreToolUse hook blocks common shell writes to corpus paths, but deliberately obscured writes are not all caught. Do not rely on the guard alone; review the staged patches yourself before running /meta-apply.
- Runs scripts: yes (Bash)
- Uses network: yes (optional MCP reviewers)
- Writes only to .aris/meta/ as intended
- Optional guard: templates/claude-hooks/corpus_write_guard.json
Tips for useful results
Let ARIS run with logging on for a while before you ask for analysis. A handful of runs gives you the skill's insufficient-data warning; a larger history gives trends you can act on.
If a previous cycle left a bottleneck log, the skill compares it with the new run and reports whether that bottleneck was resolved and where it has moved. Read the report, then stage only the patches you agree with.
- Keep logging on across several real workflow runs
- Target one skill when you want a focused change
- Check .aris/meta/bottleneck_log.jsonl to track progress across cycles
- Review every staged patch before landing it with /meta-apply
What people use it for
- Tuning default parameters such as round limits or review thresholds when users keep overriding them.
- Finding repeated failure patterns, such as timeouts, in a specific skill and proposing prompt changes.
- Adjusting the order of skills in a workflow when logs show interventions at the same stage.
- Checking whether the auto-review-loop converges quickly enough and whether its reviewer instructions need work.
Frequently asked questions
Is the meta-optimize skill free and what is its licence?
The skill is published under the MIT licence, as stated on this page. Check the licence file in the ARIS repository for the full terms.
What do I need before I can use meta-optimize?
You need the ARIS research harness installed, the meta_logging.json hook enabled in .claude/settings.json, and at least 5 logged skill invocations in .aris/meta/events.jsonl. Optional MCP reviewer backends can be used for cross-model review.
How do I install the meta-optimize skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into .claude/skills/ in a project. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
Why does meta-optimize say there is no event log?
The skill exits with an error when .aris/meta/events.jsonl does not exist. Copy templates/claude-hooks/meta_logging.json into .claude/settings.json, then run ARIS workflows so events are recorded.
Does meta-optimize change my skill files automatically?
No. It writes its report and staged patches under .aris/meta/ and does not edit skill files. A separate human-invoked skill, /meta-apply, lands approved changes.
How does meta-optimize compare with editing prompts by hand?
It works from recorded usage data, such as failure counts and score trajectories, rather than from memory of how workflows felt. It still requires your review, so it suits teams with accumulated logs more than one-off tweaks.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Meta-Optimize suits developers and teams who maintain the ARIS research harness, already run workflows with logging enabled, and have enough history to base changes on. Skip it if you have not enabled the meta_logging hook, or if you expect it to change files on its own, since landing changes is a deliberate human-run step.
Download the skill