nemotron-add-model

---
name: nemotron-add-model
description: Onboard a new model family (Nemotron or third-party) into skills/ — paper chunks, recipe summaries, context packs, and model card. Use when a contributor wants downstream skills like /nemotron-customize to be able to route to a new model.
---

# nemotron-add-model

Invocation: `/nemotron-add-model`.

You help contributors add a new model-family knowledge base to the Nemotron plugin ecosystem without getting the paper chunks, recipe summaries, context pack, or registration wrong.

## Tone

Concise. Checklist-first. Ask for missing facts before writing files.

- Status updates: ≤2 lines
- Prefer bullets and tables over long prose
- Say exactly which files you will create or change
- Do not guess model sizes, architecture labels, recipe coverage, or benchmark claims
- Always prefer the tech report HTML page over a PDF when both exist
- Never skip validation

---

## Workflow

Four phases. Always in this order.

### 1. Orient

Read these first:
- `skills/nemotron-add-step/SKILL.md`
- `skills/nemotron-nano3/SKILL.md`
- `skills/nemotron-nano3/INDEX.md`
- `skills/nemotron-nano3/paper/_overview.md`
- `skills/nemotron-nano3/recipes/overview.md`
- `skills/nemotron-nano3/context/index.toml`
- `skills/nemotron-nano3/context/quick-reference.md`
- `skills/nemotron-super3/SKILL.md`
- `skills/nemotron-super3/INDEX.md`
- `.claude-plugin/marketplace.json`

Then ask the contributor:
1. What is the model family name? (slug used in `skills/nemotron-{model}/`, for example `ultra` or `nano4`)
2. What is the tech report URL? (prefer arXiv HTML or another HTML page)
3. Does it have recipes in `src/nemotron/recipes/`?
4. What is the architecture type? (`dense`, `MoE`, `hybrid Mamba-Transformer`, or another precise label)
5. What sizes are available?
6. Which existing steps support this model? (Do any `step.toml` files need new `[[models]]` entries later?)

Use these repo conventions:
- The skill directory is `skills/nemotron-{model}/`.
- `SKILL.md` is a retrieval skill, not a code generator.
- Follow the same Locate → Retrieve → Cite pattern used by `nemotron-nano3` and `nemotron-super3`.
- `INDEX.md` is the knowledge map for the whole skill.
- `paper/*.md` files use YAML frontmatter with at least: `paper`, `model`, `section`, `paper_sections`, `title`, `summary`, `key_facts`, `related_steps`, `currency`.
- Paper chunks are question-oriented summaries of the report, not raw pasted sections.
- `recipes/*.md` files summarize the public repo path, what it reproduces, what it does not, and include `source_path` plus a `Reproduce with nemotron-customize` section.
- `context/index.toml` maps intents to the smallest useful file; `context/quick-reference.md` is the compact handoff sheet.
- `currency` is `frozen` for paper chunks and `evolving` for recipe summaries.
- Adding `[[models]]` entries to step manifests is a separate task. Do not modify step manifests here.

### 2. Generate

Create the skill directory:
- `skills/nemotron-{model}/`

Create these files:
- `skills/nemotron-{model}/SKILL.md`
- `skills/nemotron-{model}/INDEX.md`
- `skills/nemotron-{model}/model-card.md`
- `skills/nemotron-{model}/paper/` question-oriented report chunks
- `skills/nemotron-{model}/recipes/` recipe summaries if recipes exist
- `skills/nemotron-{model}/context/index.toml`
- `skills/nemotron-{model}/context/quick-reference.md`
- `.claude-plugin/marketplace.json` entry for the new skill

Generation rules:
1. Copy the live structure and tone from `nemotron-nano3` or `nemotron-super3`; do not invent a new layout.
2. Start `paper/` with `_overview.md`, then split the rest by question type: architecture, data, pretraining, SFT, RL, evaluation, safety, quantization, or another report-faithful grouping.
3. Base the paper chunks on the HTML report when available. Only fall back to PDF if no HTML source exists.
4. `model-card.md` should cover identity, released sizes/checkpoints, intended use, and headline results or deployment notes.
5. If recipes exist, add `recipes/overview.md` plus one file per public stage or major sub-stage.
6. If recipes do not exist, still create `recipes/overview.md`, but make it explicit that no reproduction recipes are available yet.
7. Each recipe summary should include:
   - `source_path`
   - what the repo exposes today
   - what it does not reproduce from the paper
   - a `## Reproduce with nemotron-customize` section
8. `context/index.toml` should include at least: identity, architecture, one training intent, one evaluation intent, and a build/customize handoff intent.
9. `context/quick-reference.md` should include model identity, sizes, architecture, public checkpoints, recipe map, and a `/nemotron-customize` step map or Explorer-mode fallback notes.
10. Register the new skill in `.claude-plugin/marketplace.json`.

### 3. Validate

Check all of these before finishing:
- Every `paper/*.md` file has valid YAML frontmatter
- Every paper chunk sets `currency: "frozen"`
- `INDEX.md` references all `paper/` and `recipes/` files
- `context/index.toml` and `context/quick-reference.md` both exist
- `recipes/overview.md` exists even if no recipes are available
- If recipe stage files exist, each one includes `source_path` and `Reproduce with nemotron-customize`
- `.claude-plugin/marketplace.json` is valid JSON

If validation fails:
1. Fix the missing file, frontmatter, or index/reference issue
2. Re-check the specific failure
3. Do not present the result until the knowledge base is internally consistent

### 4. Summarize

Show:
- What was created
- Every file added or changed
- Whether recipe summaries were created or intentionally kept minimal
- Which paper chunks were added
- The new marketplace entry name and description
- Any follow-up work deferred, such as future `[[models]]` entries in step manifests

---

## Boundaries

### Do
- Reuse the live Nano3/Super3 knowledge-base structure
- Prefer HTML report sources
- Keep paper chunks question-oriented and recipe summaries repo-oriented
- Create the context pack files every time
- Register the new skill in the marketplace metadata

### Don’t
- Don’t modify existing model skills just to normalize wording
- Don’t modify `step.toml` files here
- Don’t invent benchmark numbers, checkpoints, or unsupported recipe stages
- Don’t skip `recipes/overview.md` when recipes are missing
- Don’t blur paper claims and public recipe behavior
- Don’t use the PDF when a good HTML source exists

---

## When Stuck

- If the model slug is unclear, stop and ask what the invocation should be.
- If the report has no obvious section structure, chunk by user question type rather than mirroring section numbers mechanically.
- If recipe coverage is partial, document the real public stages and say what is missing.
- If step support is unclear, list the likely step ids in `quick-reference.md` but leave `step.toml` edits for a separate task.
- If the only source is a PDF, confirm that no HTML version exists before using it.
- After two failed validation loops, stop and report the exact file inconsistency.

---

## Calibration Examples

### Example 1 — model with recipes

User:
> Add a new model-family knowledge base for Nano4. It has an arXiv HTML report and public recipes under `src/nemotron/recipes/nano4/`.

Assistant:
> I’ll confirm the model slug, report URL, architecture type, released sizes, and which existing steps already support Nano4.
>
> Then I’ll create:
> - `skills/nemotron-nano4/SKILL.md`
> - `skills/nemotron-nano4/INDEX.md`
> - `skills/nemotron-nano4/model-card.md`
> - `skills/nemotron-nano4/paper/*.md`
> - `skills/nemotron-nano4/recipes/overview.md` plus stage summaries
> - `skills/nemotron-nano4/context/index.toml`
> - `skills/nemotron-nano4/context/quick-reference.md`
> - a marketplace entry
>
> I’ll keep paper chunks frozen, recipe summaries evolving, and I will not edit any `step.toml` files in this task.

### Example 2 — model without recipes

User:
> Add a paper-only knowledge base for Ultra. There is no public `src/nemotron/recipes/ultra/` directory yet.

Assistant:
> I’ll still build the full knowledge base shell, but `recipes/` will stay minimal.
>
> I’ll create:
> - `skills/nemotron-ultra/SKILL.md`
> - `skills/nemotron-ultra/INDEX.md`
> - `skills/nemotron-ultra/model-card.md`
> - `skills/nemotron-ultra/paper/*.md`
> - `skills/nemotron-ultra/recipes/overview.md` stating that no reproduction recipes are available yet
> - `skills/nemotron-ultra/context/index.toml`
> - `skills/nemotron-ultra/context/quick-reference.md`
> - a marketplace entry
>
> I will not invent stage summaries or pretend the paper is already reproducible from the public repo.

More agent context in NVIDIA-NeMo/Nemotron

9 other files this repository gives its agents.