Guide · Agent skill
How to use the Iterate Until Verified agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Iterate Until Verified is an agent skill, a set of instructions you add to an AI coding or assistant agent so it follows a specific working method. This guide explains what the skill changes, how to install it from the ZIP on this page, how to ask for it in natural language, and what to watch for. By the end you will be able to install the skill, trigger it on a multi-step task, and check whether the agent's evidence supports its final report.
Iterate Until VerifiedMengTo/Skills · MIT Download the skill Quick start
- 1Download the ZIP. Get the iterate-until-verified ZIP from this page and unzip it. You should see agents/openai.yaml, REFERENCES.md and SKILL.md.
- 2Choose where to install. Use ~/.claude/skills/ for all your projects or .claude/skills/ inside one project.
- 3Or upload in the Claude apps. In the Claude apps, open Settings → Capabilities → Skills and upload the skill.
- 4Make a substantial request. Describe a multi-step task and name the quality bar, such as passing tests or an independent review.
- 5Check the gate summary. Read the gate-by-gate evidence and the list of unverified items in the final response.
- 6Push back if evidence is thin. Ask the agent to rerun a specific gate or show the output behind a claim.
What the skill makes the agent do
The skill is only instructions, so it adds process rather than new tools. When it applies to a substantial task, the agent keeps your original request as the authority, turns vague quality words such as 'perfect' or 'production-ready' into pass/fail checks, splits the work into parts, and has a separate reviewer check the output against evidence before declaring it done.
It can also work in a second mode. If you ask for an improved prompt rather than the finished work, the agent returns a reusable prompt wrapper that keeps your original task unchanged inside it.
- Execute mode (default): the agent completes the task using the loop.
- Compose mode: the agent rewrites your prompt into a wrapper and does not silently start executing it.
- Verification relies on the agent actually running checks, which it can misreport or skip.
Install the skill from the ZIP
Download the ZIP for this skill from this page and unzip it. The package contains agents/openai.yaml, REFERENCES.md and SKILL.md.
For a personal install, place the unzipped skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.
You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/
- Project install: .claude/skills/ inside your project
- Claude apps: Settings → Capabilities → Skills
Trigger the skill with natural requests
You do not need a special command. The skill's description lists the situations it fits, so ask in ordinary language that matches them: fanning out work, using subagents or independent reviewers, looping until done, benchmarking against references, applying a harsh critic, comparing candidates blind, improving an existing prompt with verification, or continuing until explicit quality gates pass.
Name the quality bar you care about. Words like 'verified' or 'pass every check' point the agent toward the gate-based workflow.
- 'Use independent reviewers to check this migration plan before you report back.'
- 'Loop until the tests pass and the README matches the code.'
- 'Rewrite this prompt so it keeps my task exactly as written but adds verification steps.'
Know what the agent should produce
For substantial work, the agent should build a short acceptance matrix with a gate, a verification method, a pass condition and evidence for each. Failed gates are routed back for revision and rechecked. The agent should not finish until required gates pass or it names a concrete blocker.
Read the final report for the gate-by-gate evidence summary and the list of anything left unverified. Those two sections tell you more than a confident closing sentence.
- Each gate has a pass condition that can be observed, such as a test result or a read-back.
- Evidence can be a command, source, screenshot, output or artifact.
- Unverified items should be listed, not hidden.
Use it for substantial tasks only
The skill is designed for multi-step work that benefits from independent review. For quick, simple questions the loop adds overhead without value, so ask normally there.
The skill applies only when the agent has permission to use subagents or tools. If it lacks that permission, the skill falls back to sequential work with the same ownership boundaries.
- Good fit: a multi-file change that needs tests and an independent review.
- Poor fit: a one-line factual question.
Safety notes: scripts and network
According to the listing, this skill requires nothing beyond itself, does not run scripts and does not use the network. It is made of instruction files only.
Because the agent's tools and permissions still govern what it can do, review any external actions it proposes, such as changes to a live system, before approving them.
- Runs scripts: no
- Uses network: no
- Requires: nothing listed
What people use it for
- Reviewing a code change with a separate reviewer and a test gate before merging.
- Turning a fuzzy goal like 'production-ready landing page' into checkable criteria such as rendered sizes and accessibility checks.
- Rewriting an existing prompt into a reusable wrapper that keeps the original task unchanged.
- Comparing two drafts blind, with candidates anonymized and the same conditions applied.
- Checking a research summary against primary sources and listing contradictions found.
Frequently asked questions
How do I install the iterate-until-verified skill?
Download the ZIP from this page and unzip it. Place the folder in ~/.claude/skills/ for personal use or in a project's .claude/skills/ folder. You can also upload it in the Claude apps under Settings → Capabilities → Skills.
Does this skill require anything else to run?
The listing says it requires nothing and does not run scripts or use the network. It is instruction files only, so it adds process rather than new tools.
Is the iterate-until-verified skill free?
The skill is published under the MIT licence. The source does not list a price, so check the MIT licence text for the terms that apply to reuse.
Why isn't the skill triggering for my request?
Use wording that matches its description, such as asking for independent reviewers, looping until done, or passing explicit quality gates. Keep the request substantial, since simple questions are not a good fit.
Does the skill guarantee the work is verified?
No. It relies on the agent actually running checks, which it can misreport or skip. Ask to see the evidence behind any gate that matters to you.
How does it compare to just asking for a careful review?
The skill gives a fixed structure: explicit pass/fail gates, separate making and judging roles, and a loop that routes failures back for revision. A plain request for care does not force those steps.
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 want an agent to check its own substantial work against explicit criteria and report the evidence. Pick something else for quick questions, or for work where you need guaranteed verification, since the skill depends on the agent running and reporting its checks honestly.
Download the skill