Skip to content

Seed Your Model Card Library

Populate your model card library with cards for a shipped list of 14 frontier models in a single command.

This is the fastest way to get a working library you can browse, compare, and reference. After this tutorial you will have one card per frontier model on disk, ready for downstream tooling.

Prerequisites

  • The model-cards plugin installed — see the model-cards landing.
  • A working internet connection. The agent fetches provider docs, HuggingFace pages, arXiv papers, and web sources during research.

1. Run the seed command

/model-card seed

The command reads the shipped seed list at ${CLAUDE_PLUGIN_ROOT}/seed/frontier-models.json, prints the list and total count, and asks for one confirmation:

Research 14 cards into ~/.claude/model-cards? [y/N]

Type y to proceed. Anything else aborts.

2. Watch the per-model log

The command processes models sequentially. For each one you will see a one-line status:

[1/14] anthropic/claude-opus-4-7      created
[2/14] anthropic/claude-sonnet-4-6    created
[3/14] anthropic/claude-haiku-4-5     created
[4/14] openai/gpt-5                    skipped (refused: not confirmed)
...

Three outcomes are possible per model:

Status Meaning
created Card written to the library.
skipped (existed) A card already exists at the target path; left untouched. Use --force to overwrite.
skipped (refused: <reason>) The agent could not confirm the model exists via tier-1 (provider docs) or tier-2 (HuggingFace). No file is written. The refusal reason is logged.

Refusals are normal for models that are recent, retired, or whose provider has not published documentation under the exact model name in the seed list.

3. Review the summary

When the run completes, the command prints a four-line summary:

Created: 11
Skipped (existed): 0
Skipped (refused): 3
Failed (other): 0

The library now lives at ~/.claude/model-cards/ (or whatever MODEL_CARDS_DIR was set to). Cards are organised by provider:

~/.claude/model-cards/
├── anthropic/
│   ├── claude-opus-4-7.md
│   ├── claude-sonnet-4-6.md
│   └── claude-haiku-4-5.md
├── google/
│   ├── gemini-2.5-pro.md
│   └── gemini-2.5-flash.md
└── ...

4. Inspect a card

Open any of the generated cards. Every card has the same 10-section structure (see the card template reference) and a frontmatter sources block listing every URL the agent fetched, grouped by tier.

A typical card opens like this:

---
model_name: claude-opus-4-7
provider: Anthropic
model_version: claude-opus-4-7
last_researched: 2026-04-29
card_version: 0.1.0
researcher: model-card-researcher (claude-opus-4-7[1m])
sources:
  - tier: 1
    url: https://docs.anthropic.com/...
    fetched: 2026-04-29
  ...
---

# Model Card: Anthropic/claude-opus-4-7

## 1. Model Details

Claude Opus 4.7 is Anthropic's most capable model in the Claude 4
family [T1.1] ...

Every factual claim ends with a [T<n>.<m>] citation marker. n is the source tier (1 = provider docs, 2 = HuggingFace, 3 = arXiv, 4 = web). m is the index within that tier. Resolve the URL by looking it up in the frontmatter sources block.

5. Re-research a card

If a card came up thin (lots of "Not publicly available" sections) or the source material has changed, re-research a single card with:

/model-card create <model-name> --provider <provider>

This is the same flow as the research a model card how-to. When prompted about the existing card, choose overwrite or load-existing-as-base.

What you have now

A populated model card library on disk, with one card per confirmed frontier model. Each card carries per-claim citations and tiered source provenance. Refusals are logged but no fabricated cards exist in the library — the existence-check rule guarantees this.

Next steps