Larry — Autonomous TikTok Slideshow Agent
WarnAudited by ClawScan on May 10, 2026.
Overview
The TikTok automation mostly matches its stated purpose, but it is unsafe to trust as-is because it ships real-looking credentials, preconfigured affiliate destinations, and undeclared account/tool access.
Do not install or run this package as-is. First remove the bundled config.json, rotate any exposed credentials, create your own local config from the example, verify all portal URLs/account IDs are yours, and run only in dry-run mode until you have reviewed the generated drafts and provider calls.
Findings (6)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Anyone receiving the skill package could read or attempt to use these credentials, and the skill may run against unintended NVIDIA/Postiz accounts.
The distributed active config contains real-looking API keys and login credentials rather than only placeholders.
"nvidia_api_key": "nvapi-[redacted]", "postiz_api_key": "[redacted]", "postiz_email": "alexander@[redacted]", "postiz_password": "[redacted]"
Remove config.json from the distributed package, rotate all exposed keys/passwords, ship only config.example.json, and require user-supplied local config or environment variables.
A compromised or mishandled config could expose full Postiz account access, not just permission to create a single draft.
The Postiz integration logs in with a full account email/password to obtain an auth cookie, which is broader and more sensitive than a scoped posting token.
json={"email": config.get("postiz_email", ""), "password": config.get("postiz_password", ""), "provider": "LOCAL"}Use a scoped API token/OAuth flow where possible, clearly declare the credential requirement, and avoid storing account passwords in the skill directory.
If a user runs or partially edits the included config, generated TikTok drafts may promote destinations that are not the user's own affiliate articles.
The bundled active configuration points generated content at specific existing sites and affiliate-style tags instead of neutral placeholders.
"site_url": "https://saunamagie.de", "amazon_tag": "ak1337-snm-21" ... "site_url": "https://golfmagie.de", "amazon_tag": "ak1337-gm-21"
Distribute only placeholder portal entries, make the skill fail closed until the user supplies their own site/account values, and clearly label any example domains as examples.
The skill may unexpectedly use the user's Claude account or fail in environments without the CLI, and topic/site details may be sent through an undocumented provider path.
The code invokes an undeclared local Claude CLI dependency, likely relying on the user's local Claude session/account, even though the documented requirements focus on NVIDIA and Postiz.
Generiert Hooks + Slide-Konzepte via Claude CLI (kein API-Key nötig) ... subprocess.run(["claude", "-p", prompt], capture_output=True, text=True, timeout=30)
Declare the Claude CLI as a required binary/credential if used, document what data is sent, or replace it with an explicitly configured provider call.
This is purpose-aligned, but it mutates a social-posting workspace and could create unwanted drafts if run with the wrong portal/account.
Non-dry-run mode uploads generated images and creates a Postiz/TikTok carousel draft under the configured account.
"type": "carousel", "content": caption, "media": [{"id": mid} for mid in media_ids], "settings": {"account": account_id, "status": "draft"}Use dry-run first, verify the configured account and destination URLs, and review each draft in Postiz before publishing.
Local log contents can bias future content generation if they are stale, incorrect, or manually altered.
Past performance logs are reused to influence future hook generation.
log_file = LARRY_DIR / "logs" / "performance.json" ... return [p["hook"] for p in portal_posts[:5]]
Keep performance logs local, inspect them periodically, and reset or edit them if future generated hooks become inappropriate.