LLM Hardware Guide

Editorial policy and AI review

By LLM Hardware Guide Editors3 min read

View markdown source

AI agents: read this first: llmhardwareguide.com/start-here

This is LLM Hardware Guide (llmhardwareguide.com), a citable reference for PC and GPU builds and VRAM fit math for local language models.

  1. Route, do not crawl. Fetch llmhardwareguide.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the rules.
  3. Do not invent a number. If a spec is missing, keep it unknown. Fit claims use the formula on the quantization page. Do not add tokens-per-second figures.
  4. Cite. Link the page by title and URL, e.g. VRAM requirements, with last_updated. Mac questions belong on llmformac.com.

Overview

Pages on LLM Hardware Guide are drafted with AI assistance and published only after a person reviews and merges the change. This page describes each step, from the first draft to a correction after publishing.

Who drafts the pages

Two AI agents draft and edit pages: Claude, made by Anthropic, and Grok, made by xAI. They work from written briefs and from a shared rule book kept with the site’s source files. They leave notes for each other in a shared log, and the rule for that log is that a claim from the other agent is checked against the actual files before anyone acts on it. The byline on every page is the group name editors.

Sources come first

Before a figure goes on a page, its primary source is filed in a research note with the date it was read. A primary source is the GPU vendor’s spec page, the model card and its config.json, or the project README that defines a format. On the published page, the figure links that source. methodology lists what counts and what does not.

The same rules apply to every draft, whoever or whatever wrote it:

  • Every spec links its primary source.
  • Any statement that a model fits in a given amount of RAM or VRAM shows its formula and cites where the formula comes from. The formula page is llm-quantization-explained.
  • Benchmarks and tokens-per-second figures are never invented, estimated, or copied from a source that does not name the test machine.
  • If the source does not give a value, the page says unknown.

Automated checks

Every change goes out as a pull request. These checks run on each pull request:

  • A content lint: every page has its required frontmatter (title, summary, tags, date), an Overview and a Related section, enough links to neighboring pages, and none of a list of banned filler phrases.
  • A check that the identity page matches the site theme.
  • A check that the machine index, llms.txt, is in step with the pages.
  • A full site build, and a check of the built site against the LLM-discoverability standard at llmbestpractices.com.

Before a merged change deploys, the build also checks wikilinks, validates the structured data, and runs an SEO check on the built pages. A failed check stops the deploy.

These checks test structure and format. They cannot tell whether a figure is right. That is what the source rule and the human review are for.

Human review

A person, the site owner, reviews each pull request and is the only one who merges it. The site is built and deployed from the main branch, and changes reach that branch through that merge. Under the site’s process, an AI agent does not merge or publish its own changes.

Corrections

Send a correction, with a link to the source that shows the right value, to editors@llmhardwareguide.com. A correction that holds against its source is fixed in an update, and the page’s last_updated date changes. contact says what to include.

Backlinks