Guide · Agent skill
How to use the Meta-Optimize Harness 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 workflow project. It reads a log of past ARIS runs, finds where skills get used, fail or need human correction, and then proposes ranked changes to the instructions, default settings and review rules that ARIS uses to drive its own skills. This guide covers what the skill does, how to install it, how to trigger it, what it needs and the safety points to check before you run it.
Meta-Optimize Harness Tunerwanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Install the skill. Download the ZIP from the page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. Alternatively upload it under Settings → Capabilities → Skills in the Claude apps.
- 2Create the event log. Provide .aris/meta/events.jsonl from a Codex-compatible event logger, an external wrapper or an exported trace log, since Codex installs do not create it automatically.
- 3Check you have enough runs. Make sure the log contains at least 5 skill invocations. If not, keep using ARIS and re-run later.
- 4Ask for an analysis. Use a natural request such as "meta optimize" or "improve skills", optionally naming a target skill.
- 5Review the summary and ranked proposals. Read the findings table, the bottleneck verdict and the ranked changes before deciding anything.
- 6Apply changes yourself. Edit SKILL.md files or defaults by hand after review. The skill proposes changes rather than applying them automatically.
What is an agent skill, and what does meta-optimize do?
An agent skill is a folder containing a SKILL.md file. The file tells an AI coding agent when to use a capability and how to carry it out. Meta-optimize is a skill for maintaining the ARIS research harness: the set of skills, workflows and review loops that wrap around language models to run research tasks.
The skill looks at accumulated run history rather than at research output. It counts which skills get used, where tools fail or need manual correction, and how many review rounds it takes to reach a threshold. It then ranks proposed fixes. It is designed to propose changes, not apply them automatically, and it does not change the papers or code the workflows produce.
- Can propose edits to SKILL.md prompts, reviewer instructions and quality gates
- Can propose changes to default parameters such as retry limits and review thresholds
- Can propose changes to convergence rules and workflow ordering
- Does not optimize papers, code or experiments themselves
- Does not change MCP bridge or reviewer infrastructure configuration
Install the meta-optimize skill
Download the ZIP for this skill from the page, then unzip it into one of two locations. For a personal install available to all your projects, unzip into ~/.claude/skills/. For a single project, unzip into the project's .claude/skills/ folder. The skill's files are SKILL.md.
If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project folder
- Claude apps: Settings → Capabilities → Skills
Prepare the event log before you run it
The skill reads an event log at .aris/meta/events.jsonl. The source says Codex mirror installs do not create Claude Code hooks, and logging is not created automatically for Codex installs. You need to supply this file yourself, from a Codex-compatible event logger, an external wrapper, or a manually exported trace log.
Before running the skill, the data check looks for the file and counts skill_invoke and session_start entries. The skill needs at least 5 skill invocations. If there are fewer, it prints a warning and exits, asking you to keep using ARIS and try again later.
- File path: .aris/meta/events.jsonl
- Minimum: at least 5 skill invocations (skill_invoke events)
- If the file is missing, the skill stops with an error
EVENTS_FILE=".aris/meta/events.jsonl"
if [ ! -f "$EVENTS_FILE" ]; then
echo "ERROR: No event log found at $EVENTS_FILE"
fiHow do you trigger meta-optimize with a natural request?
The skill's description lists trigger phrases. You can ask in English or Chinese. Phrases named in the description include "meta optimize", "improve skills", "优化技能" and "分析使用记录". You can also ask for it when you want to optimize ARIS's own harness components from accumulated experience.
The skill takes an optional argument for the target, either a specific skill name or all. Keep requests specific so the agent knows the scope.
- "Run meta optimize on the auto-review-loop skill"
- "Improve skills based on the usage records in my event log"
- "分析使用记录" to analyze usage records
- "Optimize ARIS's own harness components from what we have learned"
What the skill produces
The skill writes a structured summary table of its findings. It analyses frequency (most-used skills and parameter overrides), failures (repeated error patterns such as OOM, import, compilation or timeout), convergence (average rounds to threshold and score trajectory), and human intervention (where users step in and what they correct).
It then names one current bottleneck in a single sentence and appends that verdict to .aris/meta/bottleneck_log.jsonl. The next run reads the last line first. Finally it ranks optimization targets by expected impact, with a signal, a proposed change and an expected impact for each.
- Summary table of usage, failure, convergence and intervention findings
- One-sentence bottleneck verdict appended to .aris/meta/bottleneck_log.jsonl
- Ranked table of proposed changes
Safety notes: scripts and network access
This skill runs scripts and uses the network. Its allowed tools include Bash and several optional reviewer backends (antigravity, grok and claude-review). Review the commands before you approve them, and check which reviewer backends you have configured.
The skill's rules forbid deleting safety, review or provenance steps. A model change is a reason to re-read a skill, not evidence by itself to remove anything. Treat every proposal as a draft that needs human review before it is applied.
- Bash commands run in your shell; read them before approval
- Optional reviewer MCP backends may send data over the network
- Proposals must be reviewed by a human before any SKILL.md is changed
- Safety, review, corpus/provenance and output-contract steps must not be removed
What people use it for
- A developer maintaining the ARIS harness finds that one skill fails often with timeout errors and reviews the proposed retry limit change.
- A team compares how many review rounds the auto-review-loop takes to reach its threshold and considers adjusting the threshold default.
- A researcher building their own agent workflow studies the ranked proposals as a worked example of log-driven prompt improvement.
- An engineer checks which skills users override most often, to spot defaults that might be poorly chosen.
Frequently asked questions
Is the meta-optimize skill free to use?
The source gives the licence as MIT. It does not list a price, so check the repository for any terms that apply to your use.
What do I need before I can run meta-optimize?
You need an event log at .aris/meta/events.jsonl with at least 5 skill invocations, a Bash shell for the data-check commands, and a Codex-compatible or external event logger to create the log. Optional reviewer MCP backends are referenced in its allowed tools.
Why does the skill say my log is missing?
The skill stops when .aris/meta/events.jsonl does not exist. Logging is not created automatically for Codex installs, so you must create the file from a logger, wrapper or exported trace before running it.
Does meta-optimize change my research papers or code?
No. The skill is designed to propose changes to the harness, meaning SKILL.md files, reviewer prompts and workflow defaults. It does not change the research artifacts that workflows produce.
Does meta-optimize apply its changes automatically?
No. It is designed to propose changes rather than apply them automatically, and proposals still need human review. Its rules also forbid deleting safety, review or provenance steps.
Can I use the skill with a different agent or tool?
The source describes it as Codex-compatible in terms of logging, and mentions a Claude Code-specific trigger-rate tool that has no Codex equivalent. The skill is part of the ARIS project, so check the repository for which environments are supported.
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 maintaining the ARIS research harness, teams tuning review thresholds and retry limits from evidence, and researchers who want a worked example of log-driven prompt improvement. Ordinary users who just want to write a paper or run experiments should skip it, since it needs an accumulated event log first and is about improving the tooling itself.
Download the skill