tpu_performance_autoresearch_wiki: Skill for Claude Code

.claude/skills/start-experiment/SKILL.md

start-experiment is a skill for Claude Code from vlasenkoalexey/tpu_performance_autoresearch_wiki. It costs 103 tokens per session (9,932 once invoked), scanned A, original, MIT.

A startup procedure for an autoresearch optimization loop, which repeatedly tests changes to improve a model. It prepares the selected model and execution path, then starts a supervised run on local or cluster hardware.

In plain words
What is it for?
Use it to begin model or TPU-kernel optimization sessions, choose a local chip or cluster target, record the run, and hand off to the experiment steps.
Why use it?
It removes the manual work of finding the right instructions, hardware, run target, and monitoring setup before experimentation.

Skill for Claude Code

Written for Claude Code: installed under .claude/. Also seen: reads .claude/ paths; mentions subagents; names the AskUserQuestion tool.

This is vlasenkoalexey/tpu_performance_autoresearch_wiki's own configuration. It tells Claude Code how to work on tpu_performance_autoresearch_wiki itself, so it is not a mod to install elsewhere. Copy it as a starting point and replace the rules that are about this project. Everything tpu_performance_autoresearch_wiki configures →

Reuse

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.

Copy the file
curl -O https://raw.githubusercontent.com/vlasenkoalexey/tpu_performance_autoresearch_wiki/main/.claude/skills/start-experiment/SKILL.md
Clone the repo
git clone --depth 1 https://github.com/vlasenkoalexey/tpu_performance_autoresearch_wiki

Made for: Claude Code.

Wrote 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.

agentmods badge for start-experiment

README.md
[![agentmods](https://agentmods.dev/badge/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/start-experiment/github.svg)](https://agentmods.dev/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/start-experiment)
Your own site
<a href="https://agentmods.dev/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/start-experiment"><img src="https://agentmods.dev/badge/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/start-experiment/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.

agentmods 80×15 button for start-experiment

Your own site · 80×15
<a href="https://agentmods.dev/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/start-experiment"><img src="https://agentmods.dev/badge/skills/vlasenkoalexey/tpu_performance_autoresearch_wiki/start-experiment.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 103 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 9,932 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector warn 7 Sept 2026
SkillSpector: 4 findings, up to high

These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →

  • high Privilege Escalation · line 167
    Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
    Fix: Remove references to credential paths. Use environment variables or secrets managers. For docs, use placeholder paths (e.g., /path/to/config). Never load .env or token files in production code paths.
  • high Tool Misuse · line 272
    Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
    Fix: Validate all tool parameters against an allowlist. Reject dangerous parameter values (shell=True, --force, -rf /) and use safe defaults.
  • medium Excessive Agency · line 235
    Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
    Fix: Add human-in-the-loop confirmation for destructive, irreversible, or high-impact operations. Never auto-execute commands that modify files, send data, or alter system state.
  • medium MCP Rug Pull · line 549
    Docker image references without a specific tag (:latest is implicit) or digest (@sha256:...) can be silently replaced by a malicious image.
    Fix: Pin the image: image:tag or image@sha256:abc123
How audits are shown
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5.1 $0.00103 $0.09932
Opus 5 $0.00051 $0.04966
Sonnet 5 $0.00021 $0.01986
Haiku 4.5 $0.00010 $0.00993

Measured 6d ago against content hash c73bbaa39e48, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-11, from the pricing page.

Security

Grade A, and why

start-experiment 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 6d 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.

.claude/skills/start-experiment/SKILL.md · 662 lines

How it starts

The opening of the file, as written. The whole thing — 662 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are starting one autoresearch session. Follow this sequence precisely. Do not skip steps.

Step 1 — Determine context (model, lane, parallelism)

KERNEL FAST-PATH — check this FIRST. If the invocation names a kernel (a kernel/benchmark path, a wiki/kernels/ family slug, or "tpu chip N"), this is a KERNEL session. Steps 2–8 are model-lane machinery — skip them ALL (no XPK cluster discovery, no xprof probe, no model-lane hardware question). The kernel lane asks its own, much smaller run-target question at Step 9·K below. Do exactly:

  1. Read wiki/kernel_experiments/program.md end-to-end; derive the family slug per its Quick-start rule 1.
  2. Step 9·K — choose the run target (ASK ONCE, then it is fixed for the whole run). See below.
  3. Step 9·0 — arm the watcher (kernel mode: family + home_repo from the family binding; bootstrap the binding first if new, per Quick-start rule 2).
  4. Step 9b — start marker, into the FAMILY's log: wiki/kernel_experiments/<slug>/pallas/log.md. Record the run target chosen at 9·K.
  5. Hand off: run the kernel Quick start (load /author-kernel, then K0–K9 via kexec.sh run).

Step 9·K — Run target: local chip or cluster pod (KERNEL ONLY)

The chip named in the prompt is the local default. Before arming anything, ask the user with AskUserQuestion (--yes skips the question and takes the local default — same convention as Step 8):

Question: "Run this kernel family on the local chip, or on a cluster pod?" Options:

  • Local chip <N> (recommended) — fastest edit→measure cycle; right for authoring.
  • Cluster pod — persistent GKE pod, kubectl exec. For a different target generation or more chips than local. Costs contended capacity for the whole run.

(When to pick which — the full rationale — is canonical in wiki/kernel_experiments/program.md's run-target section; don't restate it here.)

If Cluster pod:

  1. tools/kernel_exec/kexec.sh discover — prints the TPU capacity actually present (nodepool, accelerator, topology, machine type, chips). Never assume a generation; the GKE accelerator label value is generation-specific (tpu7x, tpu-v6e-slice, tpu-v5p-slice, tpu-v5-lite-podslice, …).
  2. Ask which row to use, and for how long (--hours, default 8 — the pod's hard TTL).
  3. Bring it up — it stays up for the whole run:
    tools/kernel_exec/kexec.sh up --family <slug> \
      --accelerator <A> --topology <T> [--chips N] --hours <H> --image <IMG>
    
    --image must have jax[tpu] + libtpu + kgate. --chips defaults to the node's allocatable count.
  4. kexec.sh sync --family <slug> after each K4 authoring pass, so the pod sees the current .repo.

Read the full file on GitHub · 662 lines

Changes

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.

  1. 6d ago Changed c73bbaa39e48
  2. 11d ago First seen · 662 lines · 103 tokens per session scan A 96f70bc1c648

Subscribe to this mod's changes

start-experiment is a skill published in the GitHub repository vlasenkoalexey/tpu_performance_autoresearch_wiki (55 stars, last pushed 7d ago), licensed MIT. It adds 103 tokens to every session and 9,932 once invoked, about $0.0005 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 2026-08-30.

Related

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.

Leeroo-AI/leeroopedia-mcp · 0 tokens

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.

Leeroo-AI/leeroopedia-mcp · 42 tokens

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…

AnastasiyaW/codex-claude-code-config · 96 tokens

customer_support_agent

You have access to Leeroopedia, a curated ML/AI knowledge base, via MCP tools. These are real MCP tools registered in your environment -- call them directly like any other tool. They contain framework-specific docs, code examples, API references, and best practices.

Leeroo-AI/leeroopedia-mcp · 0 tokens

self_evolve_rag

This document describes how to use the Leeroopedia MCP tools for the Self-Evolving RAG task. It is kept as a reference and is NOT included in the agent prompt.

Leeroo-AI/leeroopedia-mcp · 0 tokens

llm_post_training

You have access to the Leeroopedia MCP tools. Use them throughout this pipeline to make informed decisions. Specifically.

Leeroo-AI/leeroopedia-mcp · 0 tokens