Security audit
Security checks for vulnerabilities and agentic risk
This skill is a project-planning helper that clearly discloses web research and writing a masterplan file, with the main caveat that it writes that file without a final confirmation step.
Install only if you want a full, detailed planning workflow. Expect it to ask many project questions, use web research for current recommendations, and write the resulting plan into docs/masterplan/. Before running it in a sensitive repository, check whether an existing docs/masterplan/masterplan.md should be preserved because the skill does not require an explicit overwrite confirmation.
The README states that the skill will write a final plan to docs/masterplan/masterplan.md inside the project directory, but does not prominently warn that this creates or modifies files. Automatic write behavior without explicit user awareness or confirmation can lead to unintended repository changes, overwriting expectations, or side effects in sensitive workspaces.
The README advertises very broad natural-language triggers like 'help me plan a project' that can overlap with common user requests, increasing the chance the skill activates in situations the user did not clearly intend. In a skill that performs deep interviews, web research, and document generation, over-broad invocation can cause unexpected behavior, unnecessary data collection, or unintended file-writing workflows.
The activation criteria are extremely broad and overlap with many ordinary planning or design requests, which can cause the skill to trigger in situations the user did not clearly intend. In this skill’s context, that matters because the skill performs deep interviewing, web research, and eventually writes files, so ambiguous invocation increases the chance of unintended high-impact behavior and scope capture.
The skill explicitly instructs the agent to create directories and write files in the project without first giving a user-facing warning or obtaining confirmation. This can lead to unexpected filesystem modification, accidental overwrites, repository noise, or persistence of unwanted content, especially when combined with the skill’s broad triggering conditions.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
## 5. Tech stack (research each before presenting options)
- Frontend framework/library (WEB, MOBILE, DESKTOP, EXT) — search current recommendations for this category before offering choices
- Backend language/framework (WEB, API, AI if it has a backend)
- Database(s) — relational, document, vector (if AI/search features), cache layer — search current recommendations, don't default to what's familiar from training data without checking it's still the right current call
- Mobile: native (Swift/Kotlin) vs cross-platform (React Native/Flutter/etc.) — search current state of each before asking
- AI: which model/runtime (local model via what inference engine, or API-based), and if local — hardware constraints (RAM/GPU) the user's target machines actually have
- Hosting/infra target (cloud provider, self-hosted, serverless, on-device) — search current pricing/limits, don't assume old tier structures still apply
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
## Local AI Assistant / Agent
- Explicit resource budget (RAM/VRAM/disk) matched against the stated target hardware from the interview — a model that doesn't fit isn't a valid plan.
- Data privacy boundary stated explicitly: what, if anything, ever leaves the device, and under what condition.
- Prompt injection / tool-use safety considered if the agent can take actions (file system, shell, network) — least-privilege on what it's allowed to do without confirmation.
- Update/versioning path for the model itself, separate from the app code.
- Fallback behavior when the model is unavailable, overloaded, or produces an invalid/unsafe output.