Borrowing it
Nothing to install: this file belongs to vlasenkoalexey/tpu_performance_autoresearch_wiki. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/vlasenkoalexey/tpu_performance_autoresearch_wiki/main/.claude/skills/author-kernel/SKILL.mdgit clone --depth 1 https://github.com/vlasenkoalexey/tpu_performance_autoresearch_wikiWrote 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/vlasenkoalexey/tpu_performance_autoresearch_wiki/author-kernel)<a href="https://agentmods.dev/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/author-kernel"><img src="https://agentmods.dev/badge/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/author-kernel/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/vlasenkoalexey/tpu_performance_autoresearch_wiki/author-kernel"><img src="https://agentmods.dev/badge/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/author-kernel.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 | $0.00000 | $0.03027 |
| Opus 5 | $0.00000 | $0.01514 |
| Sonnet 5 | $0.00000 | $0.00605 |
| Haiku 4.5 | $0.00000 | $0.00303 |
Grade A, and why
author-kernel 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 10d 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 — 78 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are about to author kernel candidates for one kernel-family experiment. Read this end-to-end and keep it in scope while authoring.
The role boundary (from the Roles invariant, kernel_experiments/program.md): while authoring, you write candidates; you NEVER produce the evidence a verdict cites. Your own timing/parity numbers exist to steer iteration — label them author-side. The verdict evidence (independent re-measure, parity re-check, firing audit, capture) comes from the K6 verification step, which is independent of the author by construction: a kernel-verifier dispatch when the harness supports it, otherwise the fresh-process full-mode re-verification the process defines. This holds in solo mode too — the same agent later runs K6/K7, but authoring-phase numbers never graduate into verdict numbers. The falsification bar is public and pre-registered: you know exactly what you must beat; you just don't grade yourself.
Before authoring — the checklist (assemble it yourself in solo mode; it arrives as a brief in dispatched mode)
All of these must exist before the first candidate. Solo mode: you produced them at K1–K3. Dispatched mode: they are pasted in your brief — if one is missing (including required reading), refuse and name it; do not improvise paths.
| Item | What it is |
|---|---|
| reference module | the family binding's parity oracle (create_inputs + workload contract = op-point definition + reference semantics) — an existing reference impl in the home repo, the pinned pre-optimization version of the target, or the suite's baseline.py for benchmark families. Never edit it. |
| target path | the kernel file(s) at their real ship path in the home repo (from the family binding) — existing kernels edited in place, new kernels born at their ship path; NOT a scratch optimized.py |
| work dir | the family worktree wiki/kernel_experiments/<slug>/pallas/.repo/ on branch kernel/<slug>-vNNN (K4 creates it if missing). Everything you write goes here, nothing outside. The shared raw/code/<home_repo> checkout is the worktree PARENT only — no branch operations in it, ever; never sit on another family's branch (program.md Setup isolation invariants). |
| diagnosis | the K1 bound diagnosis: what the naive materializes, the bound, the predicted intervention class (`refute |
| falsification bar | pre-registered in the K3 stub (e.g. "p50 beats naive by >2% AND 3σ") |
parity_spec |
atol/rtol or bit-exact, from the family program.md |
| chip + lock | TPU_VISIBLE_DEVICES=<n>; wrap every device execution in flock /tmp/tpu-chip-<n>.lock (held around the run only, never around editing/compiling) |
| required reading | the kernel index's category strategy + class table/ladder, concepts/pallas-kernel.md mechanics (memory spaces, the (8,128) rule, gotcha table), BRIEFS.md earned rules, the canonical timing helper. Solo: read the pages. Dispatched: pasted content. |
activity |
optimization or capability-eval + any constraints (e.g. a no-peek list) |
| candidate plan | the enumerated plan from the K3 stub — the concrete candidate list you will execute (mechanism implementation + named tile-sweep points + named structural alternative, or a pure TUNE sweep). The falsification bar is the verdict bar; the plan is the stop rule |
Required-reading attestation: your FIRST candidate-ledger line is loaded: [<sections read>]. A ledger without it caps the experiment's verdict at inconclusive (LINT-enforced). The reading exists to prevent the compile-fix churn that eats authoring budgets — the 2026-07-11 benchmark lost 6/8 problems to authors who skipped it.
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.
- 10d ago First seen · 78 lines · 0 tokens per session scan A 9a5ebe33bddf
author-kernel is a skill published in the GitHub repository vlasenkoalexey/tpu_performance_autoresearch_wiki (55 stars, last pushed 6d ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 3,027 tokens. 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 2026-08-30.
Other skills, from other repositories
ml_inference_optimization
This document describes the Leeroopedia MCP tools available during the with-KB benchmark run. It is a standalone reference and is not fed to the agents automatically.
leeroopedia-mcp
Use Leeroopedia MCP to fetch grounded ML/AI best practices, build and review ML plans, debug failures, verify code/math correctness, and expand KB citations via getpage.
ml-research-lab
Machine-learning research loop for dataset curation, fine-tuning, evaluation, inference deployment, experiment tracking, and model explainability. Use when working on ML experiments, training data, model benchmarks, RunPod/GPU runs, classifier quality, vLLM/GGUF serving, SHAP-style model explanations, or…
experiment-audit
Use this skill for scientific and ML-research reasoning work — evaluating experimental claims, auditing training runs or ablations, checking whether a statistical claim holds up, assessing reproducibility, reconciling contradictory results, reviewing a paper's methodology or results section, writing reviewer-style…
midmem-research-tracker
Keep the repo's RESEARCH.md ledger current from the research ingestions already in the MidMem store: for every unevaluated weekly research ingestion, evaluate each paper's feasibility and applicable improvement for MidMem, and record it as Paper (cite + link) · Finding (the empirical claim that matters) · Decision…
sematyp
SemaTyP combines two data sources into a knowledge graph for drug discovery / repositioning.