AgentStack
Browse Sign in
Browse Why AgentStack Sell Docs
Sign in
SKILL unreviewed MIT Self-run

Careful

skill-manastalukdar-ai-devstudio-careful · by manastalukdar

Intercept destructive commands before execution — rm -rf, DROP TABLE, force-push, git reset --hard, and similar irreversible operations. Prompts for confirmation and suggests safer alternatives. Inspired by gstack's careful skill.

No reviews yet
0 installs
41 views
0.0% view→install

Install

$ agentstack add skill-manastalukdar-ai-devstudio-careful

Open-source listing, not yet scanned by AgentStack. Follow the source repository for install instructions.

Security review

⚠ Flagged

1 finding(s); flagged for manual review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures
  • high Destructive filesystem operation.

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

View the full security report →

Reliability & compatibility

Not yet reviewed
0 installs to date
no reviews yet
2mo ago

Declared compatibility

Claude CodeClaude Desktop

Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.

Preview Execution monitoring

We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.

How agent discovery & health will work →
Are you the author of Careful? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

Careful

I'll catch destructive commands before they run and give you a moment to confirm — or choose a safer path. Inspired by gstack's careful skill.

Token Optimization

Expected range: 100–300 tokens per intercept, 20 tokens (no destructive pattern detected)

Patterns used: Grep-based pattern matching, early exit (no match), progressive disclosure (warning → alternatives)

Early exit: If no destructive pattern is found in the proposed command, exit immediately without comment.

When to Use

Invoke before running a command you're unsure about, or configure as a pre-execution hook:

/careful rm -rf dist/
/careful git push --force
/careful "DROP TABLE users;"

Or paste a multi-line script and I'll scan the whole thing.

Step 1 — Scan for Destructive Patterns

Match the input against known high-risk patterns:

# Filesystem destruction
rm -rf | rmdir | shred | find.*-delete | find.*-exec rm

# Git irreversible operations
git push --force | git push -f | git reset --hard | git clean -fd | git rebase.*--force
git branch -D | git tag -d | git reflog expire

# Database
DROP TABLE | DROP DATABASE | TRUNCATE | DELETE FROM.*WHERE 1=1 | DELETE FROM.*no WHERE

# Process/infrastructure
kill -9 | pkill -9 | killall
docker rm -f | docker system prune | kubectl delete.*--force
terraform destroy | ansible.*state=absent

# Package/dependency
npm run clean | pip uninstall -y | pip install --force-reinstall
rm.*node_modules | rm.*__pycache__ | rm.*\.venv

# Data overwrite
dd if=.*of= | mkfs | fdisk | parted
truncate -s 0 | > file  (redirect clobber)

If no pattern matches: exit silently (the command is not flagged).

Step 2 — Risk Assessment

Rate the detected operation:

| Level | Color | Description | |---|---|---| | CRITICAL | RED | Data loss with no recovery path (DROP TABLE, rm -rf /) | | HIGH | ORANGE | Data loss recoverable with backups (git reset --hard, force push) | | MEDIUM | YELLOW | Reversible but annoying to undo (npm clean, docker prune) |

Step 3 — Warn and Offer Alternatives

⚠ CAREFUL — HIGH RISK — git push --force

This will overwrite the remote branch history. Team members with local checkouts
will have diverged histories.

Safer alternatives:
  git push --force-with-lease    ← only force if no one else pushed
  git push --force-if-includes   ← safer still (Git 2.30+)

What's your intent?
  [1] I understand the risk — run it
  [2] Use --force-with-lease instead
  [3] Cancel and explain more

Wait for user response before proceeding.

Step 4 — Alternatives Reference

File deletion:

  • rm -rf dist/trash dist/ (recoverable via Trash) or mv dist/ dist.bak/
  • find . -name "*.log" -delete → dry run first: find . -name "*.log" without -delete

Git resets:

  • git reset --hardgit stash first (recoverable), then reset if needed
  • git push --forcegit push --force-with-lease
  • git clean -fdgit clean -nfd (dry run preview)

Database:

  • DROP TABLE → rename first: ALTER TABLE users RENAME TO users_backup_20260427
  • DELETE FROM without WHERE → add LIMIT 1 first to verify filter, then remove limit

Docker:

  • docker system prunedocker system prune --filter "until=24h" (only old images)

Kubernetes:

  • kubectl deletekubectl get first to verify target, then delete with explicit name

Edge Cases

  • Pipeline commands (cmd1 && rm -rf): scan the full pipeline, flag any destructive step
  • Shell scripts: scan all lines, report all flagged lines with line numbers
  • Aliases: note that alias rm='rm -i' provides interactive prompts by default
  • CI/CD environment: if $CI=true is set, emit a machine-readable warning and exit non-zero rather than prompting
  • Already stashed/committed: if the working tree is clean before the destructive op, note that recovery is easier

Safety Note

This skill intercepts at the description layer — it does not actually block shell execution. For enforced guardrails in a CI pipeline, use a pre-commit hook or shell alias.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet, be the first.

Versions

  • v0.1.0 Imported from the upstream source.