Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add pengzhangzhi/superpowers-ml --skill ml-feedback-laddergit clone --depth 1 https://github.com/pengzhangzhi/superpowers-mlWrote 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/pengzhangzhi/superpowers-ml/ml-feedback-ladder)<a href="https://agentmods.dev/skills/pengzhangzhi/superpowers-ml/ml-feedback-ladder"><img src="https://agentmods.dev/badge/skills/pengzhangzhi/superpowers-ml/ml-feedback-ladder/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/pengzhangzhi/superpowers-ml/ml-feedback-ladder"><img src="https://agentmods.dev/badge/skills/pengzhangzhi/superpowers-ml/ml-feedback-ladder.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
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 | $0.00040 | $0.01233 |
| Opus 5 | $0.00020 | $0.00616 |
| Sonnet 5 | $0.00008 | $0.00247 |
| Haiku 4.5 | $0.00004 | $0.00123 |
Grade A, and why
ml-feedback-ladder scanned grade A with 0 findings 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 11d 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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 72 lines — stays where its author put it; the contents beside it link to each section on GitHub.
ML Feedback Ladder
Overview
In normal software, a passing test suite can mean the code works. In ML research, passing tests only show the code PATH might run - they do NOT show the method works. You need STAGED EMPIRICAL verification, ordered cheapest-to-expensive, where every cheap check GATES the expensive cluster/GPU jobs below it.
This skill OWNS the canonical ladder. Design the rungs for ONE specific experiment with your human partner before launching anything.
Core principle: Cheap checks gate expensive jobs. Never spend a slow rung to find a bug a fast rung would have caught.
Upstream: the experiment, metric, and protocol come from superpowers-ml:ml-experiment-design.
Downstream: the ladder you design here becomes verification steps in superpowers-ml:writing-plans, and the final rung hands off to superpowers-ml:ml-result-review.
The Ladder
Each rung names what it CHECKS, the ARTIFACT that proves it passed, and rough COST. Cost is relative - a rung is "expensive" if it consumes a scheduled GPU/cluster job.
| Rung | Checks | Proof artifact | Cost |
|---|---|---|---|
| R0 | Experiment card / protocol defined: question, locked primary metric, baseline, decision rule | The experiment card itself | minutes, no compute |
| R1 | Code / import / config / static sanity: it imports, config parses, paths resolve, seeds set | Clean import + config dump + linter | seconds, dev node |
| R2 | Shape / dtype / device / one-batch forward+backward: loss is finite, gradients flow | Logged shapes/dtypes/device + one non-NaN loss + non-zero grads | seconds-minutes, dev node |
| R3 | Tiny overfit: a handful of examples driven to ~zero loss (or memorized) | Loss curve collapsing to near-zero on the tiny set | minutes, dev node / 1 GPU |
| R4 | Real launcher smoke run: the ACTUAL launch path (local GPU or cluster smoke job) starts, checkpoints, logs, resumes - on tiny data/steps | Launcher exits 0, checkpoint written, logs/metrics emitted | one short job |
| R5 | Short pilot / early signal: real data, real config, truncated budget; metric is moving the right way and is stable | Early metric curve vs. baseline on the locked metric | a fraction of a full run |
| R6 | Full run / full study: the locked protocol at full budget, seeds/sweeps as specified | Complete metrics across all planned seeds/conditions | the expensive job(s) |
| R7 | Result review / decision memo: compare to baseline under the locked primary metric, decide | Decision memo (handed to superpowers-ml:ml-result-review) |
analysis time |
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.
- 11d ago First seen · 72 lines · 40 tokens per session scan A 9c10e43727ef
ml-feedback-ladder is a skill published in the GitHub repository pengzhangzhi/superpowers-ml (8 stars, last pushed 4d ago), licensed MIT. It adds 40 tokens to every session and 1,233 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
physicsnemo-discover
Official NVIDIA-authored guidance for navigating PhysicsNeMo — pick the model, datapipe, or example for a SciML/AI4Science task (surrogates, forecasting, downscaling, physics-informed, inverse, generative). Points at existing files via live repo search; never writes code. Do NOT use for installation or environment…
pysr
Use when fitting equations to data with PySR or SymbolicRegression.jl, when a user wants an interpretable formula, symbolic model, scaling law, or empirical relation discovered from numeric data, or when debugging a PySR search that is slow, stuck, or giving poor equations.
cellxgene-census-query
Query CZ CELLxGENE Census (61M+ cells). Filter by cell type/tissue/disease, retrieve expression data, and integrate with scanpy/PyTorch for population-scale single-cell analysis. Use this skill when: (1) Querying single-cell expression data by cell type, tissue, or disease, (2) Exploring available single-cell datasets…
ml-mlip-nvalchemi
GPU-accelerated batched inference for MACE, MatGL (TensorNet/M3GNet/CHGNet), and FairChem MLIPs using NValchemi, enabling parallel static, relax, and MD workflows across multiple structures simultaneously.
ml-committee-uncertainty
Quantify prediction uncertainty of MACE MLIPs using committee (ensemble) models; flag high-uncertainty structures for DFT verification.
ml-mace-finetune
Fine-tune MACE machine learning interatomic potentials on custom datasets.