Guide · Agent skill
How to use the Analyze Experiment Results agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Analyze Experiment Results skill (analyze-results) teaches an AI coding agent to read machine learning experiment output files, build a comparison table against a baseline, and write up what the numbers suggest. This guide explains what the skill makes the agent do, how to install it, how to trigger it, and what to check before you rely on its output. You do not need prior experience with agent skills to follow it.
Analyze Experiment Resultswanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Download the ZIP. Download the analyze-results ZIP from this page. It contains SKILL.md.
- 2Unzip into a skills folder. Unzip it into ~/.claude/skills/ for personal use or into your project's .claude/skills/ folder for one project.
- 3Or upload in the Claude apps. In the Claude apps, open Settings, go to Capabilities, then Skills, and upload the skill.
- 4Confirm the requirements. Make sure the agent has Bash shell access and that your project has result files in JSON or CSV format.
- 5Point the agent at your results. Ask for an analysis and name the folder, such as figures/ or results/, plus the baseline you want to compare against.
- 6Review the table and findings. Check the raw data table and each finding against your own runs before using them in a report.
What the analyze-results skill makes the agent do
An agent skill is a folder containing a SKILL.md file that gives an AI agent a set of instructions for one kind of task. When this skill is active and you ask for an analysis, the agent follows a fixed workflow: it finds result files, organizes the metrics, runs basic statistics, and writes structured findings.
The workflow has four parts. The agent first looks for JSON or CSV result files in folders such as figures/ or results/, or in other project output directories. It then builds a table with your independent variables (model type, hyperparameters, data configuration) and dependent variables (a primary metric such as perplexity, accuracy or loss, plus secondary metrics), and computes the change relative to a baseline. Next it reports mean and standard deviation when several random seeds exist, looks for trends when a parameter is swept, and flags outliers or suspicious results. Finally, for each finding it writes an observation, an interpretation, an implication and a suggested next experiment.
Install the skill from the ZIP download
Download the ZIP for this skill from this page. The skill's only file is SKILL.md. Unzip the archive so that the folder containing SKILL.md sits inside your skills directory.
For a personal install, place it in ~/.claude/skills/. For a single project, place it in the project's .claude/skills/ folder. Either way, the skill folder should contain SKILL.md.
If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project
- Claude apps: Settings → Capabilities → Skills
Check what the skill needs before you start
The skill needs two things. First, the agent needs Bash shell access so it can read and process files. Second, your project must contain experiment result files in JSON or CSV format.
The skill's metadata says it runs scripts and does not use the network. Review the allowed tools (Bash, Read, Grep, Glob, Write, Edit) before enabling it on sensitive projects, since Bash access lets the agent run commands in your environment.
- Bash shell access for the agent
- Result files in JSON or CSV format inside the project
- Runs scripts: yes. Uses network: no.
Trigger the skill with natural requests
The skill's description lists the phrases that should activate it: "analyze results", "compare", or a request to interpret experimental data. You can also pass a path or short description as an argument, since the skill takes a results-path-or-description hint.
Be specific about where the files are and what you want to know. The agent does better when it knows which metric matters.
- "Analyze results in figures/ and compare each variant to the baseline."
- "Compare the learning-rate sweep in results/ and tell me if the trend is monotonic."
- "Analyze results for the three seeds of model B and report mean and standard deviation."
- "Interpret the CSV in results/ and draft a one-sentence finding for our experiment notes."
Analyze: results/Read the output the skill produces
Every analysis should include a raw data table, a numbered list of key findings, and suggested next experiments where relevant. The skill can also propose updates to project notes or experiment reports and draft a short finding statement of one or two sentences.
Treat the findings as a structured first read. The skill reports means, standard deviations and outliers, but it does not specify a significance test. If a result will be published or used for a decision, have someone with statistics training check it.
- Raw data table
- Key findings with supporting numbers
- Suggested next experiments
- Draft finding statement for your notes
Tips for better results
Keep result files consistent. Use the same metric names and file layout across runs so the comparison table lines up. Name the baseline clearly in your request so the delta column is meaningful.
Use the skill for ML experiments. Its workflow is written around ML metrics and parameter sweeps, so it fits poorly with non-ML datasets where those concepts do not apply.
- Name the baseline run in your request
- Run several seeds if you want variation reported
- Keep metric names consistent across files
- Review the draft findings before pasting them into reports
What people use it for
- Comparing several model variants against a baseline run and seeing which metric moves.
- Reading a hyperparameter sweep to see whether a metric trends up, down, peaks in the middle, or plateaus.
- Summarising three or more random seeds as mean and standard deviation to check reproducibility.
- Drafting one-sentence finding statements for a team's experiment log.
- Spotting outlier or suspicious runs that need to be rerun before they are trusted.
Frequently asked questions
Is the analyze-results skill free?
The source material provided for this guide does not state pricing or licence terms beyond the MIT licence noted for the item. Check the repository's licence file for the full terms. The skill itself is a SKILL.md file that you install into your agent's skills folder.
How do I install an agent skill from a ZIP file?
Download the ZIP from this page and unzip it so the folder containing SKILL.md is inside ~/.claude/skills/ for personal use or the project's .claude/skills/ folder for one project. You can also upload it in the Claude apps under Settings, then Capabilities, then Skills.
What does the analyze-results skill need to run?
It needs Bash shell access so the agent can read and process files, and experiment result files in JSON or CSV format in your project. Its metadata says it runs scripts and does not use the network.
Does the skill do significance testing?
No. The skill reports mean and standard deviation and flags outliers, but it does not specify a significance test. Have someone with statistics training check results before publication.
Can I use this skill for non-machine-learning data?
It is not a good fit. The workflow is written around ML experiments, with metrics such as perplexity, accuracy or loss and parameter sweeps, so non-ML data where those concepts do not apply will not map well.
What should I do if the agent does not find my results?
Check that the files are JSON or CSV and sit in a folder the skill looks at, such as figures/ or results/, or in another project output directory. Pointing the agent at the exact path in your request helps.
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
Use this skill if you are an ML researcher or student who wants a fast, structured comparison of a batch of experiment runs, or a team that wants draft finding statements for reports. Pick something else if you need formal statistical inference or you are analysing non-ML data, because the skill does not run significance tests and its workflow assumes ML experiment conventions.
Download the skill