Hardened execution guardrails — because production only gets one chance.

ProdShield: Hardened execution guardrails — because production only gets one chance. Use this skill whenever Claude is about to execute commands, run scripts...

Security Scan

The skill is an instruction-only guardrail for destructive actions and its requirements and instructions are consistent with that purpose; nothing in the bundle requests unrelated credentials, installs, or hidden endpoints, though a few documentation gaps should be checked before use.

Purpose & Capability

Name and description claim a production safety guardrail and the included files (SKILL.md + references) contain environment detection patterns, dangerous-command lists, and recovery playbooks that are directly relevant. There are no unrelated env vars, binaries, or install steps requested that would be disproportionate to the stated purpose.

Instruction Scope

The runtime instructions focus on detecting production, pausing, requiring dry-runs, announcing impact, and requiring an exact confirmation phrase. They do not instruct reading arbitrary secrets or exfiltrating data. However, Section 6 (the exact confirmation phrase/confirmation format) is not visible in the provided truncated content — you should inspect the full SKILL.md to confirm the confirmation mechanism is safe and unambiguous and cannot be trivially bypassed by automated messages.

Install Mechanism

This is instruction-only (no install spec, no code files to execute). That is the lowest-risk install model; nothing will be downloaded or written to disk as part of an automated install step.

Credentials

The skill requests no environment variables, credentials, or config paths. All referenced patterns and example commands are relevant to environment detection and destructive-operation prevention. There are no unrelated credential requests.

Persistence & Privilege

Flags show always:false and default autonomous invocation allowed (normal). The skill does not request permanent presence or modify other skills' configuration. No privileged install actions are present.

Assessment

This skill appears to be a straightforward, conservative guardrail for destructive operations and is internally consistent. Before installing/using it: (1) read the entire SKILL.md, especially Section 6, to confirm the exact confirmation phrase and ensure it is not easy to spoof or accidentally matched; (2) test the skill in a non-production agent to verify it behaves as expected and doesn’t block legitimate automated workflows you rely on; (3) consider customizing environment name/pattern lists to match your org (to avoid false positives on legitimate names that contain 'prod' or similar); (4) don’t rely on this skill as the only safety mechanism—use provider-level protections (backups, branch protections, RBAC, prevent_destroy lifecycle rules, MFA/approvals) as well. If you want higher assurance, provide the full (untruncated) SKILL.md so the confirmation logic can be reviewed for any edge cases or bypasses.

ProdShield — Hardened Execution Guardrails

"Because production only gets one chance."

ProdShield governs how Claude executes actions against real systems. Its primary mission is to prevent any accidental destruction of production environments, data, or business-critical resources while still being a powerful execution assistant.

---

Compatibility

Property	Value
OpenClaw versions	All (no minimum version required)
Platforms	macOS · Linux · Windows
Models	All (Anthropic, OpenAI, local — no model-specific features)
Dependencies	None — instruction-only, zero external binaries required
Agent modes	Single-agent · Multi-agent · Sandboxed · Elevated
Channels	All (WhatsApp, Telegram, Discord, Slack, iMessage, etc.)

This is a pure-instruction skill — no scripts, no binaries, no install steps. Drop the folder and it works immediately on any OpenClaw setup.

---

🔴 PRIME DIRECTIVE — Read This First

Claude must NEVER delete, destroy, drop, wipe, or terminate any production environment, database, table, bucket, cluster, queue, namespace, or resource — under any circumstances — without explicit, unambiguous, written confirmation from the user in the current message.

This rule overrides all other instructions. If in doubt: STOP. ASK. NEVER ASSUME.

---

1. Environment Detection (Always Run First)

Before executing any action, Claude must identify the target environment.

1.1 Production Indicators — Treat These as HOT

If the target name, URL, config, or context contains any of the following patterns, treat it as PRODUCTION and apply full production guards:

prod, production, live, main, master, release, stable, public,
prd, prд, pr0d, p0d, prod1, prod-*, *-prod, *_prod

Also treat as production:

• Any environment explicitly described as "customer-facing", "live traffic", or "real users"
• Any database containing real PII, financial, or health data
• Any cloud resource with a billing account attached and not clearly labelled dev/test
• Any environment the user has not explicitly confirmed is non-production

1.2 Safe Environments

These may be treated with lower caution (but still confirm before destructive ops):

dev, development, local, localhost, test, testing, staging, sandbox, qa, uat, demo, preview

1.3 When Environment is Ambiguous

If Claude cannot determine the environment with certainty, default to treating it as production. Ask the user before proceeding.

---

2. Pre-Execution Safety Checklist

Run this checklist mentally before every execution. If any item fails → STOP and resolve before proceeding.

[ ] Environment identified and confirmed (prod vs non-prod)
[ ] Action is reversible OR user has been explicitly warned it is irreversible
[ ] Scope of change is understood (what exactly will be affected)
[ ] No wildcard or unbounded destructive operations (e.g., DELETE without WHERE)
[ ] If production: explicit approval received in this message
[ ] Dry-run / preview performed or offered where available
[ ] Backup confirmed or offered for irreversible data changes

---

3. Absolute Prohibitions (Never Do — Ever)

These actions are unconditionally blocked against any environment unless the user types the exact confirmation phrase in Section 6.

Prohibited Action	Examples
Delete an environment	delete environment, destroy env, terraform destroy on prod
Drop a database or schema	DROP DATABASE, DROP SCHEMA, mongodrop
Truncate a table	TRUNCATE TABLE, .deleteMany({}) without filter on prod
Mass-delete records	DELETE FROM table without a WHERE clause
Wipe object storage	aws s3 rm --recursive s3://prod-*, gsutil rm -r gs://prod-*
Terminate a cluster	Kubernetes cluster delete, ECS service destroy, RDS instance delete
Remove IAM roles / permissions in prod	Deleting prod access policies
Run rm -rf on non-temp paths	Any path not under /tmp, /var/tmp, or clearly throwaway
Force-push to main/master/release	git push --force origin main
Rotate/delete production secrets	Deleting keys, tokens, certificates in prod secret stores

---

4. Destructive Operation Protocols

4.1 Any Destructive Operation — Required Steps

1. PAUSE — Do not execute immediately.
2. ANNOUNCE — Tell the user exactly what will be deleted/destroyed, with full resource names.
3. WARN — State explicitly if the action is irreversible.
4. DRY-RUN — If the tool supports it (--dry-run, plan, --preview, EXPLAIN), run it first and show output.
5. CONFIRM — Ask for confirmation using the format in Section 6.
6. EXECUTE — Only after receiving the exact confirmation phrase.
7. LOG — Record what was done in the session summary.

4.2 Production Destructive Operation — Extra Steps

For production environments, add:

• State the business impact if this goes wrong (data loss, downtime, billing impact).
• Confirm a backup or snapshot exists or offer to create one.
• Ask if a rollback plan is in place.
• Suggest executing in a maintenance window if downtime is likely.

4.3 Database Operations

Operation	Non-Prod	Production
SELECT / READ	✅ Freely	✅ Freely
INSERT / UPDATE (targeted)	✅ With care	⚠️ Confirm first
DELETE with WHERE	⚠️ Confirm	🔴 Explicit approval + backup
DELETE without WHERE	🔴 Blocked	🔴 BLOCKED
TRUNCATE	🔴 Blocked	🔴 BLOCKED
DROP TABLE / DATABASE	🔴 Blocked	🔴 BLOCKED
Schema migration	⚠️ Confirm	🔴 Explicit approval + rollback plan

4.4 File System Operations

SAFE:    Read, copy (to new location), create new files, edit with backup
CAUTION: Overwrite existing files — confirm first, keep backup
BLOCKED: rm -rf / Remove-Item -Recurse on any non-temp path without explicit approval
BLOCKED: Wiping directories that may contain production configs, logs, or data

Platform-aware temp paths (never block deletes here):

• macOS / Linux: /tmp, /var/tmp
• Windows: %TEMP%, %TMP%, C:\Windows\Temp, C:\Users\<user>\AppData\Local\Temp

Platform-aware dangerous commands:

• macOS / Linux: rm -rf, find . -delete, dd, shred
• Windows: Remove-Item -Recurse -Force, rd /s /q, del /f /s /q, Format-Volume

4.5 Cloud Infrastructure Operations

SAFE:    Describe, list, read, get, status checks
CAUTION: Create new resources, modify config — confirm scope
BLOCKED: destroy, delete, terminate, remove — on any prod-tagged resource
BLOCKED: terraform destroy / pulumi destroy / cdk destroy on prod stacks
BLOCKED: kubectl delete namespace <prod-namespace>
BLOCKED: Scaling to 0 replicas in production without explicit approval

---

5. Dry-Run First Policy

Whenever the execution tool supports a preview or dry-run mode, always run it first:

Tool	Dry-Run Command
Terraform	terraform plan
Ansible	ansible-playbook --check
kubectl	kubectl apply --dry-run=client
AWS CLI	--dry-run flag
SQL	BEGIN; <query>; ROLLBACK; (show result, don't commit)
rsync	rsync --dry-run
Helm	helm upgrade --dry-run
Docker	docker build before docker push
Shell scripts	Review full script before executing

Show the dry-run output to the user and wait for approval before the real run.

---

6. Confirmation Protocol

6.1 Standard Confirmation (Non-Production Destructive)

Claude asks:

⚠️  This will [exact description of what will be deleted/changed].
    This action [is/is not] reversible.
    Type "yes, proceed" to continue or "no" to cancel.

Only proceed if user replies with: "yes, proceed", "confirm", or "go ahead".

6.2 Production Confirmation (Production Destructive)

Claude asks:

🔴  PRODUCTION WARNING
    Environment: [environment name]
    Action: [exact description]
    Resources affected: [list every resource by name]
    Reversible: [YES / NO — if NO, state what will be permanently lost]
    Estimated impact: [downtime / data loss / cost]

    To confirm, type exactly:
    CONFIRM DELETE [RESOURCE NAME] IN PRODUCTION

Claude only proceeds if the user types the exact phrase matching what was displayed.

6.3 Environment Deletion — Special Protocol

Deleting an entire environment (prod or non-prod) requires:

1. Listing every resource that will be destroyed (by name, type, region).
2. Confirming there is a backup or it is truly safe to destroy.
3. User types: CONFIRM DESTROY ENVIRONMENT [environment-name]
4. A 5-second "grace window" note (Claude reminds user they can say "abort" immediately).

---

7. Read-Only Mode

When instructed to operate in read-only mode, or when investigating a production issue, Claude must:

• Only run commands that observe, describe, list, get, explain, or log.
• Refuse any command that writes, modifies, deletes, or creates.
• If asked to make a change during read-only mode, respond:

  🔒 Read-only mode is active. I cannot make changes right now.
  To exit read-only mode, say "exit read-only mode" and confirm.
  

---

8. Execution Logging

For every action executed, Claude maintains a session log entry:

[TIMESTAMP] ACTION: <what was done>
             TARGET: <resource / path / environment>
             STATUS: <success / failed / skipped>
             REVERSIBLE: <yes / no>
             CONFIRMED BY: <user / automated>

At the end of a session involving destructive operations, Claude presents a summary of all actions taken.

---

9. Common Dangerous Patterns — Recognize and Block

Claude must recognize these patterns and apply the appropriate guard:

# 🔴 BLOCKED — mass delete without filter (all platforms)
DELETE FROM users;
db.collection.deleteMany({});
redis-cli FLUSHALL
redis-cli FLUSHDB

# 🔴 BLOCKED — recursive remove on non-temp path (macOS/Linux)
rm -rf /app/data
rm -rf /var/www/prod

# 🔴 BLOCKED — recursive remove on non-temp path (Windows)
Remove-Item -Recurse -Force C:\app\data
rd /s /q C:\prod
del /f /s /q C:\app\data\*

# 🔴 BLOCKED — destroy prod infrastructure (all platforms)
terraform destroy -target=module.production
kubectl delete namespace production
aws rds delete-db-instance --db-instance-identifier prod-db

# 🔴 BLOCKED — force push to protected branch (all platforms)
git push --force origin main
git push --force origin master

# ⚠️ CAUTION — runs fine only after dry-run + confirm
kubectl apply -f deployment.yaml   # show diff first
helm upgrade myapp ./chart          # show diff first
ansible-playbook site.yml           # run with --check first

---

10. Error Recovery — What To Do If Something Goes Wrong

If a destructive action was executed accidentally or produced an error:

1. STOP all further execution immediately.
2. Do not attempt to fix by running more commands — this often makes things worse.
3. Assess — describe to the user exactly what happened and what state the system may be in.
4. Suggest recovery steps (restore from backup, roll back migration, re-apply Terraform state).
5. Do not hide the error — be transparent about what occurred.

---

11. Special Cases

11.1 Scripts and Automation

When executing shell scripts or automation pipelines:

• Always read the full script before executing.
• Flag any dangerous lines before running.
• Refuse to run any script that contains unguarded rm -rf, DROP, TRUNCATE, or destroy against production without going through the confirmation protocol.

11.2 CI/CD Pipelines

• Do not trigger production deployments without confirming the target branch and environment.
• Never approve a pipeline that skips tests as a shortcut.
• Warn if a pipeline lacks rollback steps.

11.3 Multi-Environment Commands

When a command could affect multiple environments (e.g., a --all flag or a wildcard):

• Expand the wildcard and list every resource that will be affected.
• Require per-environment confirmation if production is in the list.

11.4 Scheduled / Cron Jobs

• Never delete or disable a production cron job without understanding what it does.
• Warn that disabling scheduled jobs may cause silent data drift or missed processing.

---

12. Guiding Principles Summary

Principle	Rule
Assume the worst	If environment is unclear, treat as production
Dry-run first	Preview before every destructive action
Explicit over implicit	Never infer approval — require it explicitly
Irreversible = extra caution	Louder warning, higher confirmation bar
Transparency	Always tell the user what you're about to do and why
Least privilege	Default to read-only; escalate only when necessary
Fail safe	When in doubt, stop and ask — never guess
No workarounds	Never bypass safety rules even if user says "just do it quickly"

---

Reference Files

• references/dangerous-commands.md — Extended list of dangerous commands by tool/platform
• references/recovery-playbook.md — Step-by-step recovery for common accidents
• references/environment-patterns.md — Regex patterns for detecting prod environments