safe-operations
Skill1 starsChanged 39 days agoMITView on GitHub
Rules for actions that can destroy work — writing config files, deleting, overwriting, force-pushing, running installers. Use before any operation that modifies files you did not create, before writing to a config another tool owns, before any bulk find-and-replace, and whenever an operation is hard to undo.
- Deletes or force-pushes
In this repository6 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- debug-systematicallySkill
- ralph-loopSkill
- secrets-hygieneSkill
- token-harnessSkill
- verify-before-claimingSkill
The file
We split this skill into the rules it gives an agent
10 rules. About 0 are common, 8 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 8
Every rule here comes from something that actually broke.
- Distinctive: line 32
If you cannot read it, refuse and say so.
- Distinctive: line 38
Verify afterwards that unrelated content survived — assert it, do not assume.
- Very specific: line 46
Before git checkout / reset / clean / rm -rf in a repo, run git status and stash anything uncommitted, including untracked files.
- Distinctive: line 62
Write the script to a file — a scratchpad is fine — and run the file.
- Distinctive: line 70
Prefer targeted edits over sweeping regex across many files.
- Distinctive: line 71
Never round-trip a file through a shell that guesses encodings.
- Distinctive: line 72
After any bulk edit, run the tests and grep for corruption markers.
- Very specific: line 73
Put non-ASCII in source as escapes ( ), not literal characters, so a bad round-trip cannot silently change behaviour.
- Distinctive: line 85
Remove only what you recorded installing.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill 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
Safe Operations
Every rule here comes from something that actually broke.
The one that matters most
Never treat "cannot read" as "empty".
An installer read a settings.json, JSON.parse threw on a UTF-8 BOM, the
reader returned its {} fallback, and the installer wrote that over the file.
The user's model preference and every existing hook were gone.
// Wrong: unreadable and absent collapse into the same answer
function read(p) {
try { return JSON.parse(fs.readFileSync(p, 'utf8')); } catch { return {}; }
}
// Right: absent is safe to default, unparseable is not
function read(p) {
if (!fs.existsSync(p)) return { ok: true, value: {} };
try { return { ok: true, value: JSON.parse(strip(fs.readFileSync(p, 'utf8'))) }; }
catch (e) { return { ok: false, error: e.message }; }
}If you cannot read it, refuse and say so. Do not proceed on a guess.
Before writing a file you did not create
- Back it up first. A timestamped copy costs nothing.
- Merge, never replace. Preserve every key you did not come to change.
- Verify afterwards that unrelated content survived — assert it, do not assume.
Destructive commands need a dry run
Deleting, purging and bulk-replacing should print what they would do and
require an explicit --yes to act. The cost of a confirmation is seconds; the
cost of a wrong --force is someone's work.
Before git checkout / reset / clean / rm -rf in a repo, run
git status and stash anything uncommitted, including untracked files.
Never inline a script through a shell
Nested quoting across PowerShell into node -e cost three failed attempts and a
corrupted file in one session. The layers each escape differently and the error
messages point at the wrong place.
# Wrong: quotes get eaten somewhere in the layers
node -e "const c=f.replace(/x/g,'y');..."
# Right: write the script, then run itWrite the script to a file — a scratchpad is fine — and run the file. It is readable, re-runnable, and debuggable.
Bulk find-and-replace is dangerous
A repo-wide rename here silently mojibaked every non-ASCII character, which broke a regex that matched a BOM. The pattern still looked right in review.
- Prefer targeted edits over sweeping regex across many files.
- Never round-trip a file through a shell that guesses encodings.
- After any bulk edit, run the tests and grep for corruption markers.
- Put non-ASCII in source as escapes (
), not literal characters, so a bad round-trip cannot silently change behaviour.
Idempotency
Running an installer twice must be a no-op, not a doubling. Detect prior state
structurally — a tag or key you own — never by substring-matching a
serialized blob. JSON.stringify escapes quotes, so
json.includes('path" mode') never matches and every re-run stacks a duplicate.
Uninstall
Remove only what you recorded installing. Never delete user data — memory stores, logs, scorecards — even when it lives inside your own directory.
Actions that need a human first
Publishing a package, pushing, force-pushing, posting to an external service, deleting a branch, or anything visible to other people. Prepare it fully, then ask. Approval for one action is not approval for the next.
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.

