Agent skill

Iterate Until Verified

by nirholas in nirholas/three.ws

Apply a prompt-agnostic execution and verification loop to any substantial task while preserving the original request.

Apache-2.0Auto-check passedTesting & QA

Install Iterate Until Verified

skills CLI
$ npx skills add nirholas/three.ws --skill iterate-until-verified -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install nirholas/three.ws iterate-until-verified --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/nirholas/three.ws.git skills-src && mkdir -p .claude/skills && cp -r skills-src/third_party/designcode-agent-skills/agent-skills/codex/iterate-until-verified .claude/skills/iterate-until-verified && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
iterate-until-verified
GitHub stars
227
Used in
1 other repo
Token cost
~1.9k tokens
SKILL.md length
818 words
Files
3
Skills in repo
92
Repo updated
First seen
Licence
Apache-2.0

At a glance

Apply a prompt-agnostic execution and verification loop to any substantial task while preserving the original request.

  • Works in 7 steps: Lock the original task → Convert ambition into gates → Decompose and assign → …
  • The user asks to fan out work
  • SKILL.md covers Choose the mode, 1. Lock the original task, 2. Convert ambition into gates and 3. Decompose and assign, plus 6 more sections
  • Instructions only: no scripts, shell commands, URLs or credentials in SKILL.md

What it does

Iterate Until Verified is an agent skill from nirholas/three.ws. Apply a prompt-agnostic execution and verification loop to any substantial task while preserving the original request. Use when the user asks to fan out work, use subagents or independent reviewers, loop until done, benchmark against references, apply a harsh critic, compare candidates blind, improve an existing prompt with verification, or continue until explicit quality gates pass.

Its SKILL.md is about 1.9k tokens, which your agent loads only when the skill is triggered. The skill folder holds 3 other files (for example `REFERENCES.md` and `agents/openai.yaml`).

It sits in Testing & QA, covering Quality gates. The repository describes itself as: Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web… The licence is Apache-2.0.

When your agent uses it

  • The user asks to fan out work
  • Independent reviewers
  • Loop until done
  • Benchmark against references

Example prompts

  • “/iterate-until-verified”

Workflow steps

7 steps, taken from the step headings in SKILL.md.

  1. Lock the original task
  2. Convert ambition into gates
  3. Decompose and assign
  4. Separate making from judging
  5. Match proof to the work
  6. Run the loop
  7. Stop honestly

What it can do on your machine

Read from SKILL.md and the folder at commit ddf6e49. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    No scripts in the folder and no shell commands in SKILL.md.

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Iterate Until Verified loads about 1.9k tokens when it runs. Until then it costs about 102 tokens; SKILL.md has 818 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~102
When it runs · the whole SKILL.md, loaded when a task matches
~1.9k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from nirholas/three.ws at commit ddf6e49, republished under its Apache-2.0 licence (© nirholas). 818 words, ~1,865 tokens.

Download SKILL.mdSave it as .claude/skills/iterate-until-verified/SKILL.md (or your agent's skills folder). This skill also uses 2 other files; get the full folder from GitHub.
name
iterate-until-verified
description
Apply a prompt-agnostic execution and verification loop to any substantial task while preserving the original request. Use when the user asks to fan out work, use subagents or independent reviewers, loop until done, benchmark against references, apply a harsh critic, compare candidates blind, improve an existing prompt with verification, or continue until explicit quality gates pass.

Iterate Until Verified

Preserve the task. Strengthen the process around it.

Choose the mode

  • Execute: Complete the original task with the workflow below. Use this mode by default.
  • Compose: When the user asks for an improved prompt rather than the finished work, return a reusable prompt wrapper. Keep the original task authoritative and unchanged inside the wrapper.

Do not silently switch from composing a prompt to executing it.

1. Lock the original task

Extract:

  • outcome and deliverables
  • audience and use case
  • supplied inputs and references
  • constraints, tools, formats, and exclusions
  • authorized actions and protected boundaries
  • explicit definition of done

Treat these as the task contract. Do not replace the subject, invent requirements, relax constraints, expand permissions, or let the verification method become the deliverable.

Ask a question only when a missing answer would materially change the work and cannot be discovered safely. Otherwise, state a reasonable assumption and proceed.

2. Convert ambition into gates

Translate words such as perfect, best, professional, production-ready, or AAA into observable checks. Select only the dimensions relevant to the task:

  • correctness and factual accuracy
  • completeness against the request
  • craft, clarity, and audience fit
  • usability and accessibility
  • robustness, edge cases, and regression safety
  • performance, security, or compliance
  • visual, editorial, or technical fidelity to a supplied benchmark

Create a compact acceptance matrix:

GateVerification methodPass conditionEvidence
Relevant quality dimensionTest, inspection, comparison, or read-backObservable binary conditionCommand, source, screenshot, output, or artifact

Prefer pass/fail conditions over vague scores. A strong reaction such as “wow” may be a useful signal, but it is never the only gate.

3. Decompose and assign

Split the task into the smallest meaningful workstreams with clear ownership, inputs, outputs, and integration boundaries.

  • Fan out only workstreams that are genuinely independent.
  • Keep coupled edits with one owner to avoid racing changes.
  • Give each worker the original task contract and only the context it needs.
  • Require every worker to return an artifact or evidence, not a confidence claim.
  • Keep one integrator responsible for cross-workstream consistency and regressions.

Use subagents or delegated workers when they are available, permitted, and useful. Otherwise, perform the workstreams sequentially while preserving the same ownership boundaries.

4. Separate making from judging

Do not let an implementer be the sole approver of its own work.

Give the verifier:

  • the original task contract
  • the acceptance matrix
  • the candidate artifact
  • the relevant benchmark or source material

Withhold the implementer’s rationale and self-assessment unless the verifier needs them to reproduce a check. Instruct the verifier to find failures first, cite evidence, reject unsupported claims, and return a gate-by-gate verdict.

For blind comparison:

  • anonymize and randomize candidates when practical
  • compare like with like using the same conditions
  • keep the evaluator blind to author or candidate identity, not to the task or rubric
  • do not call a comparison blind when obvious identity cues remain
Show full SKILL.md (348 more words)Show less

5. Match proof to the work

Use the strongest verification surface available:

  • Code: focused tests, typechecks, builds, linters, security checks, runtime behavior, and regression tests.
  • Visual work: rendered output at relevant sizes, interaction checks, accessibility checks, and side-by-side comparison with an accessible reference.
  • Research or analysis: primary sources, reproducible calculations, citation checks, and contradiction searches.
  • Writing: factual checks, brief coverage, audience fit, structure, and an editorial pass against representative references.
  • Plans or decisions: constraint coverage, dependency checks, failure scenarios, feasibility, and explicit tradeoffs.
  • External actions: exact target resolution followed by post-action read-back.

Never substitute a self-rating for evidence. Never invent a benchmark, source, test result, screenshot, or blind verdict.

6. Run the loop

Repeat:

  1. Produce or improve the candidate.
  2. Run every applicable gate.
  3. Record pass, fail, or blocked with evidence.
  4. Route each failure to the responsible workstream.
  5. Make the smallest revision that addresses the evidence.
  6. Re-run the failed gate and any affected regression gates.
  7. Integrate only verified work.

Continue while required gates fail and a safe, in-scope action can make meaningful progress. Do not churn on the same approach after repeated failure; change the approach or report the blocker.

7. Stop honestly

Finish only when:

  • every required gate passes
  • the integrated result still satisfies the original task
  • regressions relevant to the changed work have been checked
  • evidence supports the final claims
  • remaining unknowns are disclosed

Stop as blocked when a required gate depends on missing access, unavailable inputs, new authority, or an infeasible constraint. Name the exact blocker and the minimum next action. Do not weaken a gate merely to declare success.

Compose mode template

When returning an enhanced prompt, use this shape:

text
Use an iterative execution-and-verification workflow around the authoritative task below.

AUTHORITATIVE TASK
<preserve the user's original task here without changing its subject, deliverables, or constraints>

PROCESS
1. Extract the task contract and convert subjective quality language into observable acceptance gates.
2. Decompose independent workstreams and fan them out when delegation is useful and permitted.
3. Keep one integrator responsible for consistency.
4. Assign an independent verifier that sees the task, rubric, candidate, and references—but not the implementer's self-assessment.
5. Verify with task-appropriate evidence. Use anonymized side-by-side comparison when a real comparable benchmark exists.
6. Route failed gates back to the responsible workstream, revise, and re-check affected regressions.
7. Do not finish until every required gate passes or a concrete blocker is proven.

FINAL RESPONSE
Return the deliverable, a concise gate-by-gate evidence summary, and anything still unverified. Do not claim checks that were not run.

Adapt the process to the task. Do not copy domain-specific tools, benchmarks, or quality claims from another prompt unless they apply here.

Completion checks

  • The original task remains authoritative.
  • Subjective ambition became observable gates.
  • Independent work was separated without creating racing edits.
  • Making and judging were assigned to different roles.
  • Benchmarks were real, comparable, and honestly labeled.
  • Failed gates drove revisions.
  • The final claims match the collected evidence.

© nirholas, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 2 other files in third_party/designcode-agent-skills/agent-skills/codex/iterate-until-verified of nirholas/three.ws.

  • SKILL.md
  • REFERENCES.md
  • agents/openai.yaml

Open the folder on GitHubat commit ddf6e49

Used in 1 other repository

We found 1 copy of this SKILL.md (exact, near-identical or edited) in other folders, from 1 other GitHub owner. This page covers the copy in nirholas/three.ws, which our catalogue first saw on October 7, 2026.

Compare with similar skills

Iterate Until Verified next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Iterate Until Verified compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Iterate Until Verified this skillnirholas/three.ws2271 repos~1.9kAutomated safety check: PassApache-2.0
Prompt Generatorcatlog22/Claude-Code-Workflow2.1k1 repos~4.7kAutomated safety check: NotesMIT
Dev ReviewFHIR/fhir-codegen154—~5kAutomated safety check: PassMIT
Suede MCP Release QAJasonColapietro/suede-creator-skills127—~2.1kAutomated safety check: PassMIT
Flow Parallelnyldn/claude-octopus4.2k1 repos~7.8kAutomated safety check: PassMIT
Multi Agent Task Orchestratorsickn33/agentic-awesome-skills47k2 repos~1.5kAutomated safety check: PassMIT

Similar skills

  • Prompt Generator

    catlog22/Claude-Code-Workflow

    Generate or convert Claude Code prompt files — command orchestrators, skill files, agent role definitions, or style conversion of existing files.

    2.1k GitHub starsUsed in 1 repo~4.7k tokens
    Testing & QAAuto-check: notes
  • Dev Review

    FHIR/fhir-codegen

    Performs a two-track code-quality and QA review in the roles of a staff-level Engineering Lead and QA Lead, then synthesizes both critiques into a single analysis.md.

    154 GitHub stars~5k tokensUpdated today
    Testing & QAAuto-check passed
  • Suede MCP Release QA

    JasonColapietro/suede-creator-skills

    Checks a Suede AI MCP server release against a live process: the full JSON-RPC lifecycle, schemas, annotations, malformed input, catalog agreement and install docs.

    127 GitHub stars~2.1k tokensUpdated today
    Testing & QAAuto-check passed
  • Flow Parallel

    nyldn/claude-octopus

    Decompose and execute large changes, migrations, or multi-issue fixes in parallel with quality gates

    4.2k GitHub starsUsed in 1 repo~7.8k tokens
    Testing & QAAuto-check passed
  • Multi Agent Task Orchestrator

    sickn33/agentic-awesome-skills

    Route tasks to specialized AI agents with anti-duplication, quality gates, and 30-minute heartbeat monitoring

    47k GitHub starsUsed in 2 repos~1.5k tokens
    Testing & QAAuto-check passed
  • Pcb Design

    oaslananka/kicad-mcp-pro

    Safe KiCad PCB design assistance workflow using KiCad MCP board inspection, placement, routing, stackup, and quality-gate tools.

    120 GitHub stars~1.5k tokensUpdated today
    Testing & QAAuto-check passed

More from nirholas/three.ws

All 92 skills in this repo
  • Add Shader Cursor Trail

    nirholas/three.ws

    Add the Shaders WebGPU mouse effect used for the Tidal Commons hero: a white twinkling halftone cursor trail driven by ChromaFlow, masked through a DotGrid, finished with chromatic ripples and film…

    227 GitHub starsUsed in 1 repo~760 tokens
    Auto-check passed
  • Publish Project To GitHub

    nirholas/three.ws

    Package a finished local project into an intentional GitHub repository, create a strong README and visual preview, push it safely, configure a public GitHub Pages URL when the project is compatible…

    227 GitHub starsUsed in 1 repo~2.2k tokens
    Auto-check: notes
  • Audit a website or digital experience against its supplied source references for originality and plagiarism risk.

    227 GitHub starsUsed in 1 repo~1.8k tokens
    Auto-check passed
  • Turn a completed daily UI inspiration capture into exactly five original landing-page builds, one per separate Codex task, using Sites.

    227 GitHub starsUsed in 1 repo~1.9k tokens
    Auto-check passed
  • Write Like Meng On X

    nirholas/three.ws

    Write, rewrite, review, or continuously refine X/Twitter posts in Meng To's current voice using his deduplicated authored-post corpus, personal and product context, shared resources, and Content…

    227 GitHub starsUsed in 1 repo~2k tokens
    Auto-check passed
  • Browser Video Recording

    nirholas/three.ws

    Create polished 60 fps 4:3 4K browser screen-recording style videos from Codex in-app browser captures, with browser-only crop, natural macOS cursor styling, deliberate click choreography…

    227 GitHub starsUsed in 1 repo~1.5k tokens
    Auto-check passed

Questions about Iterate Until Verified

What does Iterate Until Verified do?

Apply a prompt-agnostic execution and verification loop to any substantial task while preserving the original request. ws. Apply a prompt-agnostic execution and verification loop to any substantial task while preserving the original request.

When should I use Iterate Until Verified?

Iterate Until Verified fits situations like: the user asks to fan out work; independent reviewers; loop until done; benchmark against references.

How do I install Iterate Until Verified in Claude Code?

Run `npx skills add nirholas/three.ws --skill iterate-until-verified -a claude-code`. Or copy the skill folder (third_party/designcode-agent-skills/agent-skills/codex/iterate-until-verified in nirholas/three.ws) into .claude/skills/iterate-until-verified in your project. Claude Code loads it when a task matches its description.

How do I install Iterate Until Verified in Codex?

Run `npx skills add nirholas/three.ws --skill iterate-until-verified -a codex`. Or copy the skill folder (third_party/designcode-agent-skills/agent-skills/codex/iterate-until-verified in nirholas/three.ws) into .agents/skills/iterate-until-verified in your project. Codex loads it when a task matches its description.

Can I use Iterate Until Verified in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add nirholas/three.ws --skill iterate-until-verified -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/iterate-until-verified, .gemini/skills/iterate-until-verified, .github/skills/iterate-until-verified and .opencode/skills/iterate-until-verified in your project.

What does Iterate Until Verified need to run?

SKILL.md names no scripts, command-line tools or credentials: Iterate Until Verified is instructions for the agent only.

Does Iterate Until Verified access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Iterate Until Verified safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Iterate Until Verified use?

Iterate Until Verified is published under the Apache-2.0 licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Iterate Until Verified use?

About 1.9k tokens (SKILL.md is roughly 7.5k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Iterate Until Verified?

Skills that share tags, products or a category with Iterate Until Verified: Prompt Generator (catlog22/Claude-Code-Workflow, 2.1k stars), Dev Review (FHIR/fhir-codegen, 154 stars), Suede MCP Release QA (JasonColapietro/suede-creator-skills, 127 stars) and Flow Parallel (nyldn/claude-octopus, 4.2k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Iterate Until Verified?

nirholas (a GitHub user) maintains it in nirholas/three.ws, which has 227 GitHub stars. The repository holds 92 skills in this directory. The repository was last updated on October 8, 2026.

Source: nirholas/three.ws on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.