AGENTS.md prompt generator

What this AGENTS.md prompt generator actually creates

This page produces a request you can paste into Claude or ChatGPT when you want help drafting, reviewing, or migrating a repository instruction file. It does not produce a secretly generated AGENTS.md behind the form. It does not call a model, open your repository, judge the quality of an existing file, or write anything to disk. Every input that affects the prompt is visible: target platform, task, receiver, scope, preset, repository-access assumption, verified facts, and optional existing instruction text.

That boundary is useful because repository instructions should be evidence, not plausible boilerplate. An install command that merely resembles a common Node project can waste time or change a lockfile. A deployment rule copied from another repository can be worse. The Node and TypeScript, Python, monorepo, generic-project, and global presets therefore expose categories of questions; they never fill those questions with commands, paths, frameworks, or policies. Leave an unknown blank and the prompt tells the receiving assistant to inspect evidence or ask.

AGENTS.md and CLAUDE.md are different targets

Codex and Claude Code load instructions through different file names and documented hierarchies. The official Codex AGENTS.md guide describes a global file under the configured CODEX_HOME, with ~/.codex as the default, then project guidance discovered from the root toward the current working directory. A nearer file can add more specific guidance. In the same directory, an AGENTS.override.md file takes precedence over AGENTS.md. The open AGENTS.md format is ordinary Markdown and does not impose mandatory headings.

The official Claude Code memory guide documents user, project, project .claude, and project-local CLAUDE.md locations. It also documents @path imports. Claude Code does not load AGENTS.md directly, so one supported migration pattern is a short CLAUDE.md that imports @AGENTS.md and adds only verified Claude-specific material. The migration mode exposes that strategy explicitly rather than pretending the two platform formats have identical loading semantics.

Target, receiver, task, and scope stay separate

The target answers “which instruction system will consume the eventual file?” The receiver answers “where will I paste this request?” A ChatGPT prompt can ask for a Claude Code file, and a Claude prompt can ask for a Codex file. Switching receivers changes only the request’s address; it does not switch platform facts, filename, or scope. This separation prevents a common category mistake where “Claude prompt” is treated as proof that the desired artifact must be CLAUDE.md.

Create asks for a complete proposed file plus evidence gaps. Audit request asks the receiver to inspect the supplied evidence, name conflicts and stale or unverifiable rules, and return a proposal separately; the generator never says that audit already happened. Migration request requires a visible source format and strategy. An independent-file migration asks for explicit mappings and omissions, while the Claude import strategy is available only for the documented AGENTS.md-to-CLAUDE.md direction.

Existing instructions are untrusted input data

You can paste text or open a local .md or .txt file. File decoding is fatal UTF-8: malformed bytes are rejected instead of becoming replacement characters, and invalid intake does not erase text already in the form. A UTF-8 byte-order mark is removed, CRLF or CR newlines are normalized to LF, and the local receipt names that normalization. The file never leaves the browser.

The generator serializes supplied facts and existing instructions as canonical one-line JSON inside a clearly labeled UNTRUSTED_INPUT_DATA block. Embedded newlines and fence characters are JSON-escaped, so a copied instruction such as “ignore the request above” remains data for the receiving assistant to inspect rather than becoming an invisible generator rule. This is a prompt-injection boundary, not a claim that a later model is perfectly secure.

Deterministic output and a dated fact ledger

Pressing Generate lazily loads a local worker. The same canonical input and guidance snapshot produce the same LF-only prompt, one final newline, target path, byte count, line count, and fingerprint. Editing any field marks the visible receipt stale and disables copy and download until you generate again. Long output is clipped only in the on-page preview; copy and download retain the complete prompt. Clipboard success appears only after the browser resolves the write, and a rejected write leaves download available.

Platform behavior changes, so the prompt says “snapshot reviewed” with the date rather than “latest.” The source ledger used here was reviewed on August 14, 2026 and is scheduled for review by October 13, 2026. A static-build assertion blocks this route after that date until the official links are checked and the ledger is deliberately renewed. This does not prove the sources stay unchanged between reviews; it makes the freshness assumption visible and prevents an old snapshot from being released indefinitely by accident.

Review the prompt before asking for a file

Read the target path and truth boundary first. Verify destructive commands, credential handling, release permissions, generated folders, and deployment language against repository evidence. If the receiving assistant can inspect the repository, require it to cite the files or command output behind every project-specific rule. If it cannot, keep missing facts as questions. After you receive a proposed file, compare it with any current version using Markdown diff, edit it in the Markdown editor, and run the repository’s real checks before writing or committing it.

Global instruction prompts deserve extra restraint because they affect unrelated projects. The global preset excludes project commands and architecture sections by design. A project preset cannot be selected for a global target, and a global preset cannot be used to disguise an under-specified project prompt. The safest instruction file is not the longest one; it is the one whose scope, evidence, permissions, and exceptions can be understood by a person before an assistant acts.