stack
AGENTS.md2 starsChanged 4 days agoMITView on GitHub
Workspace skills live in .agents/skills/ and are shared through a-novel-kit/stack. Publish skill changes on a feature branch with a pull request. Use these repository copies in this environment; keep personal skill installations outside this workflow. Never touch what you do not own unless the developer explicitly says to. You own only what your current session created: its worktrees, branches, stacks, containers, issues and pull requests. A bare "clean up" or "clear", of worktrees, issues or anything else, stays inside that…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository40 other files it gives its agents.
- .agents/skills/AGENTS.mdAGENTS.md
- attribute-ai-commitsSkill
- check-accessSkill
- choose-dependencySkill
- coordinate-landingSkill
- develop-featureSkill
- document-codeSkill
- git-conventionsSkill
- implement-featureSkill
- manage-versionsSkill
- monitor-ciSkill
- open-pull-requestSkill
- plan-client-server-boundarySkill
- plan-featureSkill
- plan-ui-designSkill
- prefer-small-solutionsSkill
- prepare-releaseSkill
- register-ai-agentSkill
- resolve-pr-feedbackSkill
- triage-issuesSkill
- use-a-novel-cliSkill
- use-discordSkill
- write-bash-scriptsSkill
- write-design-systemSkill
- write-dockerfilesSkill
- write-frontendSkill
- write-frontend-testsSkill
- write-github-actionsSkill
- write-go-kitSkill
- write-go-serviceSkill
- write-goSkill
- write-go-testsSkill
- write-infraSkill
- write-js-packageSkill
- write-openapiSkill
- write-platformSkill
- write-project-docsSkill
- write-protoSkill
- write-sqlSkill
- write-svelteSkill
The file
We split this AGENTS.md into the rules it gives an agent
8 rules. About 2 are common, 6 are distinctive, and 0 are very specific to this project.
≈2
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 7
Never touch what you do not own unless the developer explicitly says to.
- Distinctive: line 12
Remove what you own by exact name, never with a store-wide prune, and state that rule in every prompt that delegates work able to create or …
- Distinctive: line 21
When the developer asks for a command, your next reply is that command and nothing else.
- Common: line 27
When a step needs the developer's own hands, such as a secret, a web console setting or an approval, hand it over ready to act on: the …
- Common: line 33
Before an operation they may lack, such as an approval, a release, a privileged label, a repository or org setting, a secret, Google Cloud …
- Distinctive: line 45
Write no em dash (—) in anything you produce: interface text, documentation, comments, commits, issues, pull requests and replies.
- Distinctive: line 50
Prefer the smallest complete solution: fewer instructions for a reader to hold, fewer moving parts, and clear, idiomatic control flow.
- Distinctive: line 55
Before planning, writing, changing, or reviewing code, load prefer-small-solutions.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Working principles
Workspace skills live in .agents/skills/ and are shared through a-novel-kit/stack.
Publish skill changes on a feature branch with a pull request. Use these repository copies
in this environment; keep personal skill installations outside this workflow.
Never touch what you do not own unless the developer explicitly says to. You own only what your current session created: its worktrees, branches, stacks, containers, issues and pull requests. A bare "clean up" or "clear", of worktrees, issues or anything else, stays inside that set; a global cleanup that owns the whole workspace is always requested explicitly. Another session's checkout can be clean, merged and idle while that session still runs in it, so when ownership is unclear, ask. Remove what you own by exact name, never with a store-wide prune, and state that rule in every prompt that delegates work able to create or remove anything. See workspace hygiene.
Open issues and pull requests only in the a-novel and a-novel-kit organizations. Never open one
on any other repository, even to report or fix a defect you found there: work around it in the
workspace, link an existing upstream report when there is one, and leave filing a new one to the
developer. Forking or cloning an outside repository to test something stays fine.
When the developer asks for a command, your next reply is that command and nothing else. Read only what its arguments need, then hand it over. Leave out the procedure around it, the follow-up commands and the checks it suggests: the developer is usually acting on a clock, as during a planned downtime, and each extra step delays them. Verify and monitor after the hand-off, and report only what they ask for.
When a step needs the developer's own hands, such as a secret, a web console setting or an approval, hand it over ready to act on: the exact command with every value you know filled in, a direct link to the runbook section that covers it, or both when the step mixes a console and a command. Never only name the step.
You act with the developer's access, and most developers hold less than these skills describe. Before an operation they may lack, such as an approval, a release, a privileged label, a repository or org setting, a secret, Google Cloud or Discord, check their access with check-access. When they lack it, say what is missing and who holds it, then stop: never look for a workaround. A command they ask for but cannot run gets that answer instead of the command.
Commands handed to the developer to run must paste cleanly into zsh with oh-my-zsh, which bash
accepts too. Put a space before a ;, & or | that follows a URL: url-quote-magic escapes it
on paste, so …/pull/12; done arrives as …/pull/12\; done and fails to parse. Quote arguments
holding [], ? or *, list loop items literally, and quote every expansion: zsh neither splits
an unquoted $VAR into words nor passes an unmatched glob through.
Write no em dash (—) in anything you produce: interface text, documentation, comments, commits,
issues, pull requests and replies. It spends width for nothing and reads as machine-written. Use a
period, a colon, a comma or parentheses instead. A page title separates the page from the brand
with |, and an empty table cell stays empty.
Prefer the smallest complete solution: fewer instructions for a reader to hold, fewer moving parts, and clear, idiomatic control flow. Preserve required behavior and the project's architecture, dependency policy, security, and testing standards. Simplify the whole affected path before shortening individual functions; do not compress formatting or hide complexity to reduce line count.
Before planning, writing, changing, or reviewing code, load prefer-small-solutions. Keep its decision rule active through implementation and final review. For software development, also load develop-feature to choose the current stage: local draft, issue and draft-PR review, approved-scope testing, or final cleanup. Load other skills according to the work they govern.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

