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 topprismdata/cultivating-ml-agent --skill onnx-minimal-network-designgit clone --depth 1 https://github.com/topprismdata/cultivating-ml-agentWrote 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/topprismdata/cultivating-ml-agent/onnx-minimal-network-design)<a href="https://agentmods.dev/skills/topprismdata/cultivating-ml-agent/onnx-minimal-network-design"><img src="https://agentmods.dev/badge/skills/topprismdata/cultivating-ml-agent/onnx-minimal-network-design/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/topprismdata/cultivating-ml-agent/onnx-minimal-network-design"><img src="https://agentmods.dev/badge/skills/topprismdata/cultivating-ml-agent/onnx-minimal-network-design.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 | $0.00144 | $0.00982 |
| Opus 5 | $0.00072 | $0.00491 |
| Sonnet 5 | $0.00029 | $0.00196 |
| Haiku 4.5 | $0.00014 | $0.00098 |
Grade A, and why
onnx-minimal-network-design 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 12d 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 — 95 lines — stays where its author put it; the contents beside it link to each section on GitHub.
ONNX Minimal Network Design
Problem
Neural network golf competitions require the smallest possible ONNX network
that correctly implements a transformation. Score = max(1, 25 - ln(memory + params)).
Smaller networks → exponentially higher scores. The key insight: memory dominates cost (intermediate tensor bytes), not parameters.
Solution
Core Scoring Formula
memory = Σ(bytes of all intermediate tensors) # input/output are FREE
params = Σ(initializer element counts)
score = max(1, 25 - ln(memory + params))
Design Principle 1: Single-Node = Zero Memory
A single-node graph (1 Conv/Gather/Transpose) has zero intermediate memory (output tensor is free). Only params count.
Implication: A 1-node Conv with 900 params (cost=900, score=18.2) beats a 4-node net with 200 params + 5000 memory (cost=5200, score=16.3).
Design Principle 2: Conv is a Linear Classifier
A Conv node maps 3×3 one-hot neighborhoods → output channels. This is a linear classifier on 90-dim features (9 positions × 10 colors).
To find Conv weights for a local rule:
- Collect all (3×3 neighborhood, output color) pairs from training data
- Verify the rule is deterministic (same neighborhood → same output)
- Train LogisticRegression (multi-class, C=1e4, solver='newton-cg') on the pairs
- Extract coef_ as Conv W[10,10,3,3], intercept_ as B[10]
- Verify ONNX runtime output matches ground truth on ALL examples
Limitation: If LogReg accuracy < 100%, the rule is NOT linearly separable. Need nonlinear ops (Cast/Equal/Where) — a single Conv cannot implement it.
Design Principle 3: Gather for Color Permutations
A pixel-wise color map (e.g., swap color 3↔4) is a 1-node Gather:
- idx = [0,5,6,4,3,1,2,7,9,8] (10-element lookup table)
- Gather(input, idx, axis=1) maps each one-hot vector to the permuted one
- Cost: 10 params, 0 memory → score ≈ 22.7
Design Principle 4: Sparse Initializers REJECTED
Converting dense sparse weights to sparse_initializer reduces params but
onnx.checker.check_model(full_check=True) rejects sparse Conv weights:
W typestr: T, has unsupported type: sparse_tensor(float)
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.
- 12d ago First seen · 95 lines · 144 tokens per session scan A 1c45f9c64651
onnx-minimal-network-design is a skill published in the GitHub repository topprismdata/cultivating-ml-agent (5 stars, last pushed 14d ago), licensed MIT. It adds 144 tokens to every session and 982 once invoked, about $0.0007 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
huggingface-hub
Hugging Face Hub CLI (hf) — search, download, and upload models and datasets, manage repos, query datasets with SQL, deploy inference endpoints, manage Spaces and buckets.
tensorboard
Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit.
mlflow
Track ML experiments, manage model registry with versioning, deploy models to production, and reproduce experiments with MLflow - framework-agnostic ML lifecycle platform.
datachain-knowledge
Use whenever datasets, cloud storage buckets, or data pipelines are mentioned — creating, saving, querying, listing, exploring, deleting, or processing data in S3, GCS, Azure Blob, or local storage. Also use when running any script that may create datasets as a side effect. Maintains a knowledge base at dc-knowledge/…
prompt-scanner
A scanner for text sent to an AI agent, looking for prompt injection and jailbreak attempts. Prompt injection is text that tries to override an agent's instructions; a jailbreak tries to bypass its safety limits.
install-openviking-memory
Install and configure the OpenViking long-term memory plugin for OpenClaw via natural conversation. Once installed, the plugin automatically captures facts from chats and recalls relevant context before each reply (auto-capture + auto-recall, cross-session). Covers prerequisites, install through OpenClaw's plugin…