Free · Runs in your browser
agents.md generator
Fill in a short form and get a root AGENTS.md your coding agent can actually follow: stack, verify commands, hard constraints, and a session continuity protocol. Nothing you type leaves this tab.
AGENTS.md preview
37 lines
# AGENTS.md ## Working rules - Keep changes scoped to what was asked. Do not refactor adjacent code on the way past. - Read the nearby files before writing. Match the patterns already in this repo over the ones you prefer. - Prefer the smallest change that fully solves the task. - If a requirement is ambiguous, state the assumption in your reply rather than guessing silently. ## Continuity **At session start** - Read this file before touching anything. - Read the most recent decisions or notes in the repo and treat them as current. - Do not assume state from an earlier session survived. Re-check it. **At session end** - Record what was verified by running a command, separately from what was assumed. - Record anything left half-done and where it stopped. - Never claim done without having run the Verify commands in this session. ## Handoff End every session with this block so the next session starts from fact, not memory. ``` Files touched: Commands run (and their result): Verified: Assumed, not verified: Open risks: ``` <!-- Generated with the agents.md generator at https://night-forge.com/tools/agents-md-generator -->
How to use it
- Fill in the fields, or press Load sample to see the shape first.
- Paste your real verify commands, exactly as you would type them in a terminal.
- Copy or download the file and commit it at your repository root as AGENTS.md.
- Tell your agent to read it at the start of every session, then check the handoff block at the end.
Why this matters
Most agent failures are not reasoning failures. They are memory failures. The agent solved the problem yesterday, the context window closed, and today it starts from nothing: re-reading the same files, re-deciding the same trade-offs, and quietly undoing a fix it made itself.
A stack list does not fix that. A read-write protocol does. The generated file tells the agent what to read before it touches anything, and what to write down before it stops: what was verified by running a command, what was only assumed, and where the work stopped. That is the difference between a file the agent skims and a file that carries state across the gap.
The Verify section does the other half. It turns "done" into something a checklist can settle, which is why it is the section worth being pedantic about. One file covers one repository well. Keeping continuity across many sessions, many repos, and a context reset in the middle of a task is a larger routine, and that is what the Session Continuity bundle below is.
Questions
- What is AGENTS.md?
- A plain Markdown file that tells a coding agent how to work in your repository: what the stack is, which commands prove a change works, and what it must never do. Agents read it the way a new contributor reads a README, except they read it every session.
- Where does the file go?
- The repository root, next to package.json, committed to version control. Root placement is the whole point: an agent that has to hunt for the file will not find it, and a file that is not committed does not exist for the next session.
- AGENTS.md or CLAUDE.md?
- AGENTS.md is the vendor-neutral name that several agent tools now look for. CLAUDE.md is Claude Code's own convention. If you use one agent, use the name it reads. If you use more than one, keep AGENTS.md as the source of truth and have the tool-specific file point at it rather than duplicating the rules.
- Why put verify commands in the file?
- Because 'done' is the word agents get wrong most often. Listing the exact lint, typecheck, test, and build commands turns done into something checkable: the agent either ran them in this session or it did not, and the checklist makes the difference visible in the reply.
- Does this replace a continuity system?
- No. The generated file gives you the read-write protocol and a handoff block, which is enough for one repository and one agent. A continuity system covers the parts a single file cannot: durable session state, decision records that survive a context reset, and a routine for restarting work you did not finish.
Keep the context between sessions
The generated file gets one repository started. The bundle is the full routine: session state that survives a reset, decision records the agent has to read, and a restart procedure for work that stopped halfway.
Session Continuity Config Bundle, $31Want the config layer too? Claude Code Master Config Bundle, $37.60