Getting it into your agent
This one installs as part of its plugin. Adding the marketplace and installing the plugin brings it with everything else the plugin ships.
/plugin marketplace add infiniV/ultra-ml-intern/plugin install ml-internWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/infiniv/ultra-ml-intern/ml-intern)<a href="https://agentmods.dev/skills/infiniv/ultra-ml-intern/ml-intern"><img src="https://agentmods.dev/badge/skills/infiniv/ultra-ml-intern/ml-intern/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/infiniv/ultra-ml-intern/ml-intern"><img src="https://agentmods.dev/badge/skills/infiniv/ultra-ml-intern/ml-intern.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00131 | $0.02134 |
| Opus 5 | $0.00066 | $0.01067 |
| Sonnet 5 | $0.00026 | $0.00427 |
| Haiku 4.5 | $0.00013 | $0.00213 |
Grade B, and why
ml-intern scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Reads MCP configurationmediumAgent snooping
mcp.json carries server URLs and auth tokens; reading it lets a mod discover and abuse other integrations.
- **MCP server** — Hugging Face MCP at `https://huggingface.co/mcp`, declared in `.mcp.json`. Activates when the user has `HF_TOKEN` set; otherwise everything still works via `WebFetch` + the bundled shell helpers. How it starts
The opening of the file, as written. The whole thing — 148 lines — stays where its author put it; the contents beside it link to each section on GitHub.
ML Intern — paper research
You are an ML research assistant. Your job is to answer methodology questions from published papers you actually read, not from training-time recall. Every claim you report is attributed to a specific paper, section, and number.
Core principles
Your memory of ML papers and HF library APIs is outdated and lossy. Citation counts, SOTA claims, dataset names, hyperparameters, and TRL/Transformers API shapes all drift. You will confidently produce a plausible recipe that no paper published. Read first, then write.
Read the paper, don't summarize a summary. paper_text.py writes a paper's
real text to disk and you Read it. WebFetch answers a prompt using a small
summarizer model, so its output is a paraphrase — quotes taken from it can't be
verified and its section numbers are guesses. Use WebFetch for repos, docs,
and Hub pages; never for papers.
Report what the evidence supports. A section that must be filled gets filled, and what fills it is invention. "No contradictions surfaced" and "the paper doesn't state a learning rate" are findings. A short report backed by real quotes beats a long one padded to look thorough.
Skip research only for trivial non-code questions.
The research loop
- Find the landmark paper(s) for the task or domain.
- Crawl the citation graph for recent downstream work — see
references/paper-crawl.md. - Read the most promising papers — fetch each to disk with
paper_text.py, then read the method, experiments, and results. Recent + high-citation + strong benchmarks first. Abstracts lie by omission. - Extract the recipe: dataset, training method, hyperparameters that produced the published result. Attribute every claim to a specific result (e.g. "Dataset X + method Y → 85.3% on benchmark Z, §4.2 Table 3").
- Validate the recipe against reality — does the dataset exist on Hub? Does the base model? Are the columns what the method needs?
- Report with working URLs: arXiv, Hub, and the paper's official repo.
What ships with it
12 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- references/paper-crawl.md 11 KB
- references/ultra-research.md 10 KB
- scripts/crawl_arxiv.sh 6.8 KB runs code
- scripts/download_paper.sh 4.4 KB runs code
- scripts/hf_paper_meta.sh 4.6 KB runs code
- scripts/inspect_dataset.sh 3.2 KB runs code
- scripts/merge_papers.sh 2.7 KB runs code
- scripts/paper_text.py 9.0 KB runs code
- scripts/recommend_papers.sh 1.5 KB runs code
- scripts/research_slug.sh 1.1 KB runs code
- scripts/snippet_search.sh 3.2 KB runs code
- scripts/verify_quotes.py 6.0 KB runs code
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 148 lines · 131 tokens per session scan B d324320ac5aa
ml-intern is a skill published in the GitHub repository infiniV/ultra-ml-intern (3 stars, last pushed 22d ago), licensed MIT. It adds 131 tokens to every session and 2,134 once invoked, about $0.0007 per session on Opus 5. A static security scan graded it B with 1 finding (reads mcp configuration). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
prompt-engineering
Create, update, review, or discuss an LLM prompt — a system prompt, a skill, or an agent. Use when writing or improving a prompt, discussing a skill or agent, diagnosing prompt failures, or when the user says a prompt needs work.
review-prompt
Review LLM prompts against the prompt-engineering skill's principles — leading with where each line came from — and report the findings without modifying files. Use when reviewing prompt quality, auditing a prompt, evaluating a system prompt, or checking whether prompt issues are high-confidence and fixable.
ccc-prompt-fix
Fix and sharpen a prompt. Diagnoses it against the 6 prompt-quality patterns, returns a tightened rewrite with the reasoning, and suggests the right library prompt for your task.
eval-set
Build and run a project-specific retrieval eval so changes to your rules or base prompt are scored, not eyeballed. Mirrors tests/groundtruth.json + clawness eval: you write prompt→expected-rule cases, then measure MRR@k and hit-rate before and after an edit. Run it after trimming a base prompt into ranked retrieval…
oracle-grill-me
Grill the user relentlessly about a plan, decision, or idea, maintaining the project's domain model (CONTEXT.md, ADRs) as decisions crystallise. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.
reinforcement-learning
Reinforcement Learning best practices for Python using modern libraries (Stable-Baselines3, RLlib, Gymnasium). Use when: Implementing RL algorithms (PPO, SAC, DQN, TD3, A2C) Creating custom Gymnasium environments Training, debugging, or evaluating RL agents Setting up hyperparameter tuning for RL Deploying RL models…