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 OutlineDriven/outline-driven-development --skill cudagit clone --depth 1 https://github.com/OutlineDriven/outline-driven-developmentWrote 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/outlinedriven/outline-driven-development/cuda)<a href="https://agentmods.dev/skills/outlinedriven/outline-driven-development/cuda"><img src="https://agentmods.dev/badge/skills/outlinedriven/outline-driven-development/cuda.svg" alt="Measured on agentmods" 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 | $0.00046 | $0.01810 |
| Opus 5 | $0.00023 | $0.00905 |
| Sonnet 5 | $0.00009 | $0.00362 |
| Haiku 4.5 | $0.00005 | $0.00181 |
Grade A, and why
cuda 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 yesterday.
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 — 117 lines — stays where its author put it; the contents beside it link to each section on GitHub.
CUDA
Contract
| Field | Bound contract |
|---|---|
| Trigger | CUDA C or C++ kernel code is being written, optimized, or reviewed: launch configuration, memory hierarchy, streams, nvcc flags, Thrust, or known performance traps such as warp divergence and uncoalesced access. |
| Authority | Read-only. The skill emits guidance and code; it writes no files. Nothing to roll back. No remote mutation. |
| Side effect | Guidance, code samples, and flag recommendations in chat output. |
| Done | A launch configuration, memory plan, build command, and pitfall review exist for the kernel in question, each grounded in the toolkit version the project pins. |
Inputs
- Kernel goal (required): the operation to move to the GPU, or the existing kernel under review.
- GPU model (required if code must run on it): compute capability from
nvidia-smi --query-gpu=compute_cap. - Toolkit version (required if not inferrable):
nvcc --version. The grounded current stable is CUDA Toolkit 13.3 Update 1. - Build system (optional): existing nvcc or CMake CUDA flags.
Procedure
- Identify the GPU compute capability and toolkit version. CUDA 13.x supports Turing (sm_75) and later; older targets require an older toolkit. Done when: the
sm_XXtarget and toolkit line are recorded, or the mismatch is reported. - Write or review the launch: threads grouped into blocks, blocks into a grid. Cap a block at 1024 threads. Compute
blocks = (n + threads - 1) / threadsand guard every load and store withif (i < n). Check each launch withcudaGetLastError()and each sync with a second check. Done when: the launch math covers every element and both error checks are present. - Place data at the right level of the hierarchy. Registers hold per-thread scalars. Shared memory holds tiles reused across the block. Global memory holds large arrays, read coalesced. Constant memory holds broadcast read-only data. Latency numbers below are order-of-magnitude guides on current parts, not constants:
What ships with it
1 file 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.
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.
- yesterday First seen · 117 lines · 46 tokens per session scan A ea71bdb484c3
cuda is a skill published in the GitHub repository OutlineDriven/outline-driven-development (52 stars, last pushed yesterday), licensed Apache-2.0. It adds 46 tokens to every session and 1,810 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 2026-09-06.
Other skills, from other repositories
embedded-stm32
Best practices for embedded C/C++ development on STM32 microcontrollers using the HAL, covering peripherals, DMA, interrupts, memory constraints, and hardware-focused testing. Use when writing STM32 HAL code, configuring peripherals generated by STM32CubeMX, working with interrupts or DMA, debugging with SWD/JTAG…
acad-arx-wizard
Agentic ObjectARX project scaffolding for AutoCAD 2027 / Visual Studio 2026. Replaces the broken .vsz VsWizardEngine wizard with a PowerShell script that generates identical C++ project files. Works for new ARX/DBX/CRX projects and add-on class wizards (Jig, Reactors, Custom Object, MFC, .NET Wrapper, COM Wrapper…
hls-c2rtl
Translate C / C++ / SystemC algorithmic descriptions into synthesizable RTL via High-Level Synthesis. Use when the user says "HLS", "C to RTL", "C++ to Verilog", "Vitis HLS", "Catapult", "XLS", "algorithmic design", or provides a reference implementation in software and asks for hardware.
stm32-baremetal
STM32 bare-metal skill for CMSIS-only MCU projects. Use when scaffolding STM32 firmware without HAL, configuring clocks/RCC, using CMSIS headers, or building with arm-none-eabi-gcc. Activates on queries about STM32 bare metal, CMSIS without HAL, STM32F4/H7 bring-up, or minimal Makefile CMake for Cortex-M.
mspm0-skill
Tool-neutral CLI agent rules for TI MSPM0 development with SysConfig, DriverLib, CCS, Keil/uVision, CMake/GCC/OpenOCD, and supported board references. Use when an agent needs to inspect or modify MSPM0 projects, validate SysConfig output, package examples, or work on NUEDC embedded firmware.
rocm
AMD ROCm GPU computing stack for HIP kernel development and GPU-accelerated library usage. Use when: writing HIP kernels (.hip files), using rocBLAS/rocFFT/rocRAND/rocSOLVER/rocSPARSE/hipBLAS/hipBLASLt/hipTensor/hipSPARSELt/rocALUTION compute libraries, profiling with rocProfiler or rocprof, porting CUDA code to HIP…