Editorial Rubric

How We Verify a Claude Build

A claim that something was built with Claude is not enough. The gallery asks what was built, who built it, how Claude Code was involved, what evidence is public, and what remains unverified.

Updated 2026-09-199 minverificationeditorial policyClaude Code evidence

Key Facts

Page summary for citation

  • Editorial Rubric: A claim that something was built with Claude is not enough. The gallery asks what was built, who built it, how Claude Code was involved, what evidence is public, and what remains unverified.
  • Evidence trail: 6 sources, including Anthropic Docs, Anthropic Docs, Anthropic Docs.
  • Internal context: related pages include Gallery, Open Source, About.
Anthropic Docs6verification testsIdentity, artifact, Claude surface, model generation, evidence, and limits.Anthropic DocsPrimarysource preferenceOfficial docs, customer stories, public repositories, and builder posts come before summaries.

"Follows your standards"

Claude Code GitHub Actions docs - source

The six-question test

Every gallery candidate is reviewed against six questions. If the public record cannot answer them, the build can still be mentioned as a lead, but it should not be treated as verified.

  • Who is the builder or deploying organization?
  • What artifact exists: app, workflow, template, agent, integration, or case study?
  • Which Claude surface is involved: Claude Code, Claude Code on the web, Claude Agent SDK, Claude in Chrome, Claude Cowork, the Claude API, Claude.ai, or managed agents?
  • Which model generation does the source name, if any: Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, or nothing at all?
  • What public evidence supports the claim?
  • What are the limits: private code, self-reported metrics, missing benchmarks, or unclear deployment status?

Naming the model without overclaiming

Most published Claude builds never name a model. That is not a disqualifier, but it does constrain what the page can say. A case study from an earlier product era describes the tool as it was then, and its metrics cannot be quietly relabeled as Opus 5 or Fable 5.1 results just because those are the current models.

So the rule here is narrow. If the source names the model generation, the build card names it too. If the source does not, the card says the source does not. Anthropic also documents different default effort levels per surface, which means even a correctly named model can behave differently in Claude Code than in Cowork, and a build note that omits the surface has omitted part of the result.

  • Name the model only when the primary source names it.
  • Name the surface always, because it changes both effort defaults and the evidence trail.
  • Date the claim, so a reader can tell which product era a metric belongs to.

Evidence tiers

Official customer stories are strong evidence for organizational adoption, but the numbers are still reported by the source, not independently audited here. Public repositories are strong evidence for structure and implementation, but not necessarily for production use.

Builder posts can be valuable when they include concrete details, screenshots, performance claims, repository links, or product URLs. Secondary coverage is useful for discovery, but it should not replace the primary source.

Why the rubric matters

Agent-built software invites sloppy claims because the phrase "built with AI" can mean anything from autocomplete help to autonomous multi-hour changes. The site uses specific language so readers can compare examples honestly.

That is why page copy uses verbs like reports, describes, states, and says when the source is self-reported. The goal is not cynicism. The goal is durable, citable usefulness.

FAQ

Verification questions

Does a project need public source code to qualify?

No. Enterprise case studies can qualify if the source is official and specific. Public source code is preferred for templates and community tooling.

Do you include Claude API products?

Only when the page is about a build pattern adjacent to Claude Code or Claude Agent SDK. Pure Claude API app coverage belongs elsewhere unless it teaches a Claude Code build pattern.

Can a maintainer submit a project?

Yes. The build should include a public URL, repository or writeup, Claude involvement, license or access details, and evidence of what works today.

Why not publish tutorials here?

The domain lane is project examples and build evidence. Step-by-step Claude Code practice belongs on the tutorial sibling, claudeshipscode.com.

Which Claude model built these projects?

Usually the source does not say. Where it does, the build card names the model generation. The current generally available lineup is Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5, but older case studies predate those models and their metrics should not be attributed to them.

Do builds made in Claude Cowork or Claude in Chrome count?

Yes, when the artifact and evidence are public. Those surfaces leave documents and step logs rather than commits and pull requests, so the build note has to say which evidence trail exists before quoting any outcome.

Primary Sources

Source trail

Links are cited for factual claims on this page. Access dates are kept so future updates can re-check drift.

Claude Code overview

Anthropic Docs - official docs - accessed 2026-07-06

Open citation

Current capability overview: local, web, IDE, git, MCP, hooks, skills, and agents.

Choosing a Claude model

Anthropic Docs - official docs - accessed 2026-09-19

Open citation

Current generally available lineup and official guidance on picking between Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5.

Claude Cowork

Anthropic - official product page - accessed 2026-09-19

Open citation

Agentic surface for non-coding knowledge work, described as using the same agentic approach as Claude Code, with scheduled and unattended runs.