Guide · Agent skill
How to use the Result-to-Claim Gate agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The result-to-claim skill is an agent skill for research projects that helps you decide what your finished experiments actually show. It gathers results from your project, sends them to a secondary Codex reviewer agent, and returns a judgment on which claims are supported, which are not, and what evidence is missing. After this guide you will know how to install the skill, how to ask for it in plain language, and how to read its output before you commit to conclusions in a paper or rebuttal.
Result-to-Claim Gatewanshuiyin/Auto-claude-code-research-in-sleep · MIT Download the skill Quick start
- 1Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
- 2Confirm the reviewer backend. Make sure Codex or another configured reviewer backend is available, as the skill calls a secondary reviewer agent for its judgment.
- 3Finish the experiments. Run the main experiments first. This skill is for main results, not sanity checks or early debugging.
- 4Ask for the check in plain language. Name the intended claim and the source of the results, such as a W&B run or EXPERIMENT_LOG.md.
- 5Let the agent gather results. The agent reads W&B history, your experiment logs, the tracker and the research contract, then assembles deltas, baselines and caveats.
- 6Review the verdict and route. Check claim_supported, the missing evidence list and the confidence level, then follow the route the agent recommends.
What is an agent skill, and what does this one do?
An agent skill is a folder containing a SKILL.md file that tells an AI agent how to carry out a specific task. When a request matches the skill's description, the agent loads those instructions and follows them.
This skill is meant to run after experiments finish and before you write the paper or start ablations. The agent collects results from sources such as W&B, EXPERIMENT_LOG.md, EXPERIMENT_TRACKER.md, training log files over SSH, and idea-stage/docs/research_contract.md. It then runs a deterministic check that the cited values exist in the files, and sends the results and intended claim to a secondary Codex agent for judgment.
- Judges which claims the numbers support, which they do not, and what evidence is still missing.
- Routes to a next action: pivot, run supplementary experiments, or confirm the claim.
- Writes the verdict into your project notes, such as findings.md, when a claim is partial, unsupported or confirmed.
Install the skill from the ZIP or the Claude apps
Download the ZIP for this skill from this page and unzip it. The skill is a single file, SKILL.md. Put the unzipped folder in one of two places depending on how you want to use it.
For all your projects, place it in ~/.claude/skills/. For one project only, place it in that project's .claude/skills/ folder.
If you use the Claude apps instead, open Settings, go to Capabilities, and upload the skill under Skills.
- Personal install: ~/.claude/skills/result-to-claim/
- Project install: .claude/skills/result-to-claim/ inside your project folder
- Claude apps: Settings → Capabilities → Skills → upload
Check the requirements before you start
The skill expects a configured reviewer backend. Its examples use Codex, and the file refers to a spawn_agent call with a Codex model. Other reviewers can be configured through the shared reviewer settings, including optional MCP reviewer servers named antigravity, grok and claude-review.
The deterministic check needs Python 3 and the optional evidence_check.py script. If you use W&B as a result source, you need a Weights & Biases account with API access. If you read remote training logs, you need SSH access to that machine.
- Codex or another configured reviewer backend
- Python 3 for the optional evidence_check.py script
- Weights & Biases account and API access, if W&B is a source
- SSH access, if reading remote training logs
- Claude tools used: Bash, Read, Grep, Glob, Write, Edit, plus optional MCP reviewer servers
Trigger the skill with a natural request
You do not need a special command. Describe the situation in plain language after your experiments have finished. The skill's description covers requests about judging what results support, what they do not, and what evidence is missing.
Give the agent the intended claim and where the results live. The skill accepts an optional argument describing the experiment or W&B run, so including the run ID helps. Good requests name the claim you are testing and the source of the numbers.
- "My ablation runs are done. Do the results support the claim that the attention module helps? Check the W&B run for the main table."
- "Experiments on the new loss finished. Before I start the paper, judge what the numbers support and what is still missing."
- "Use result-to-claim on the run in EXPERIMENT_TRACKER.md marked DONE and tell me whether to pivot, add experiments, or confirm."
Read the output and act on the route
The reviewer returns claim_supported as yes, partial or no, with a confidence level. The agent then follows the route for that verdict. A no verdict records a postmortem and asks you to pick a new idea or approach. A partial verdict revises the working claim to what is supported, records the gap, and calls for supplementary experiments followed by a re-run. A yes verdict records the confirmed claim and checks whether ablations are still incomplete.
If an EXPERIMENT_AUDIT.json file exists, the skill also reads its integrity_status. A fail status adds an integrity concern and lowers confidence to low. If no audit exists, the verdict is labelled provisional, and the pipeline continues.
What people use it for
- Deciding whether main results justify the claim in a paper's abstract before drafting it.
- Checking an ambiguous ablation outcome for an independent second opinion before a rebuttal.
- Confirming that a team's experiment tracker and cited values match before committing to conclusions.
- Deciding whether to narrow a claim to one dataset or run a supplementary experiment.
Frequently asked questions
How do I install the result-to-claim skill?
Download the ZIP from this page and unzip it. Place the folder in ~/.claude/skills/ for personal use, or in your project's .claude/skills/ folder for one project. In the Claude apps, upload it under Settings → Capabilities → Skills.
Is the result-to-claim skill free?
This page does not state pricing or licence terms beyond the skill's MIT licence. The reviewer backend, such as Codex, and any W&B account are separate services you must have access to.
What do I need to run result-to-claim?
You need a configured reviewer backend such as Codex, Python 3 for the optional evidence_check.py script, and a Weights & Biases account if you use W&B as a source. SSH access is needed only if you read remote training logs.
Does the skill run scripts or use the network?
Yes. It runs scripts, including an optional Python evidence check, and it uses the network to reach the reviewer backend, W&B and any remote logs you point it at. Review the files before you run it in sensitive environments.
Can I use this skill for a quick sanity check?
The skill is designed for main results. Sanity checks and early debugging are not a good fit, because the reviewer call adds cost and delay.
Why does the verdict say same-family or provisional?
The Codex reviewer is the same model family as the agent by default, so the file marks its judgment as provisional. A stronger, independent judgment needs a reviewer from a different model family.
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 when your main experiments are finished and you need a structured check of what the results support before writing a paper or rebuttal. Skip it for sanity checks, early debugging, projects without logs or a stated claim, and cases where you need a fully independent judgment, since the default Codex review is same-family and provisional.
Download the skill