discoveryfiles.ai
Project archive

For AI agents working on optional files

Start with the practical guide and a concrete purpose. The original eleven-file suite is legacy experimental material; do not install it by default.

Authoring guidance v1.2.5 · Updated 2026-09-13

Why this page exists

Use this procedure when an agent prepares an optional file: identify the source for each fact, resolve missing information and check the result.

For what a valid file must contain, the specification documents describe the legacy proposal only. For new llms.txt work, use the current upstream proposal.

Current references

Use the current guide for new work. These public resources are separate from repository instructions for coding tools.

The prime directive

Never invent a fact about the target business. Use confirmed information from the owner, the maintained public website or reliable primary records for the same entity. Resolve contradictions before publishing; omit unsupported optional facts.

Optional summaries create another place to maintain facts. Incorrect information can be copied into other outputs; publishing a file does not make it authoritative or guarantee that an AI system uses it.

Omit unsupported optional fields. Ask for missing facts needed to complete the chosen task.

What to point an agent at

Three resource types, all public, no key required. For example, ask an assistant to read https://discoveryfiles.ai/agent/authoring.md, then follow it.

/agent/authoring.md text/markdown

The full authoring procedure in Markdown. Point an agent here first.

/agent/manifest.json application/json

The authoring procedure and rules as JSON, with metadata and template URLs for the archived catalogue.

/agent/templates/{filename} text/plain

Each template served raw, so an agent never has to scrape the HTML page to get one.

fetch the manifest 
curl -sS https://discoveryfiles.ai/agent/manifest.json

Where facts may come from

Check each value against a public source or information the owner supplied. Use the categories below to decide what to copy, confirm or omit.

Use verified public sources

Check the source, entity and currency of each fact. Public availability does not resolve conflicting information or authorize publishing private details.

  • Services and products described on the site
  • Public contact addresses, phone numbers and postal addresses
  • Office or location names listed publicly
  • Social profile URLs linked from the site
  • Existing page URLs for the link sections of llms.txt
  • The site’s primary language
  • Legal name and registration details in an official register, matched to the business
  • Published prices with their currency, scope and applicable date

Ask the site owner

Use existing instructions and verified sources first. Ask only about unresolved facts, preferences or permissions needed for the task. Omit unsupported optional fields.

  • Conflicting legal identities or registration details that sources do not resolve
  • Service exclusions or regions not served when the public scope is unclear
  • Unclear price validity or responsibility for maintaining copied prices
  • Training permissions not already authorized by the owner
  • Unresolved preferences for brand names or public contact addresses

Never write

Do not publish unsupported claims or information the owner has not authorized for public use.

  • Unsupported promotional claims such as "leading" or "best-in-class"
  • Unverified or outdated prices, or copied prices with no way to keep them current
  • Headcount, revenue or funding figures without a reliable primary source
  • Comparative claims without dated evidence using comparable measures
  • Testimonials, reviews or quotes you cannot source
  • Credentials, API keys, internal hostnames or staging URLs
  • A permissive AI training grant the owner has not explicitly given

The procedure

  1. 01

    Choose a concrete purpose

    Read the practical guide at https://discoveryfiles.ai/guide and identify the requested outcome and intended consumer. Do not install the legacy eleven-file suite by default; llms.txt is optional.

  2. 02

    Read the existing site

    Inspect the relevant public pages and existing files before editing. Use confirmed owner information, maintained pages and reliable primary records for the same entity. Preserve maintained content within the requested scope.

  3. 03

    Resolve only relevant gaps

    Use information already supplied by the owner. Ask only for missing facts necessary to the chosen task; omit unsupported optional fields.

  4. 04

    Follow the chosen format

    For an optional llms.txt implementation, consult https://llmstxt.org/ and the intended consumer documentation. Our archived templates contain project-specific requirements and are not a current upstream specification.

  5. 05

    Check the facts and links

    Keep summaries consistent with their public source pages. Do not assume a separate identity file is authoritative when facts conflict; resolve the conflict against confirmed information.

  6. 06

    Keep crawler policy explicit

    Use robots.txt and provider-supported controls for the owner’s intended search, training and retrieval policy. Our experimental robots-ai.txt is not an established control. Do not change permissions without authorization.

  7. 07

    Verify and report the limits

    Check actual response bodies, HTTP status, content types and links. Validate the selected format. Report sources, edits and omissions. A valid or fetched file is not evidence of indexing, citation or influence on an answer.

Authoring rules

This list is also served in the manifest and Markdown procedure.

rules 
- Never invent facts, quotes, credentials or private business information.
- Use the current practical guide to scope new work; do not install the legacy file suite by default.
- Treat the legacy catalogue priorities and required fields as historical proposal metadata only.
- Use llms.txt only for a concrete optional workflow; follow the current upstream proposal and intended consumer.
- Keep published summaries aligned with confirmed facts on the public website.
- Include prices or comparisons only when relevant, verified and maintainable; preserve their source, date and scope.
- Do not grant training permissions or alter crawler policy without owner authorization.
- Do not claim these files control AI answers or improve visibility without evidence specific to the consumer and outcome.
- Date provider claims using an actual source review; distinguish published rules, observed retrieval, indexing and training.
- Distinguish registered offices, customer-facing locations and service areas; do not invent branches, reviews or business-profile eligibility.
- Treat AGENTS.md, CLAUDE.md and GEMINI.md as instructions for supporting project tools, not a universal public search instruction channel.

Legacy template URLs

These are original templates for maintaining the legacy proposal. Every value that must be supplied is wrapped in [SQUARE BRACKETS]. Check for unresolved placeholders before publishing.

File Original priority Template URL
llms.txt Required /agent/templates/llms.txt
llm.txt Required /agent/templates/llm.txt
llms-full.txt Conditional /agent/templates/llms-full.txt
llms.html Recommended /agent/templates/llms.html
identity.json Required /agent/templates/identity.json
ai.json Recommended /agent/templates/ai.json
ai.txt Recommended /agent/templates/ai.txt
brand.txt Recommended /agent/templates/brand.txt
robots-ai.txt Optional /agent/templates/robots-ai.txt
faq-ai.txt Recommended /agent/templates/faq-ai.txt
developer-ai.txt Conditional /agent/templates/developer-ai.txt

Check your own work for leftovers

Before handing files back, grep them. A published file still containing [Official Business Name] is unfinished content. 

grep -l "\[" llms.txt ai.txt brand.txt faq-ai.txt developer-ai.txt robots-ai.txt

This authoring guidance is published by discoveryfiles.ai. The file definitions it refers to live in the project repository under an MIT licence. See the license and attribution notes.