Skip to content
Back to skills

Project Context

ASecurity

Provides project context (tech stack, conventions, key rules) to subagents. Not a user command - loaded by the orchestrator when dispatching review subagents.

  • 7 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added September 23, 2026
ai-agentsjavascripttypescriptpythonrustgojavarubyreactnodeexpress

Security analysis

A100/100

Scanned September 23, 2026

npx -y skills add mayankmankhand/llm-peer-review --skill project-context --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Project Context?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Project Context
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/mayankmankhand-project-context/badge)](https://www.skillsdirectory.com/skills/mayankmankhand-project-context)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: project-context
description: Provides project context (tech stack, conventions, key rules) to subagents. Not a user command - loaded by the orchestrator when dispatching review subagents.
user-invocable: false
allowed-tools:
  - Read
  - Glob
  - Grep
---

# Project Context

Gather lightweight project context so subagents can make informed decisions. This skill is not user-invocable - it runs automatically when the orchestrator dispatches review subagents.

## What to Detect

<procedure>

1. **Project type** - Check for these markers in order:
   - `package.json` - Node.js / JavaScript / TypeScript
   - `requirements.txt` or `pyproject.toml` - Python
   - `go.mod` - Go
   - `Cargo.toml` - Rust
   - `Gemfile` - Ruby
   - If multiple markers exist, note that (e.g., a Node.js project with Python scripts)

2. **Project rules** - Read these files if they exist:
   - `CLAUDE.md` in the project root
   - `.claude/rules/*.md` for toolkit-level rules
   - Look for critical rules that subagents must follow (e.g., the auto loop in `.claude/skills/shared/hitl-loop.md` and its always-ask actions (M9)). Note: a per-run "report only" opt-out (M10) is conversational, never on disk - the orchestrator's dispatch prompt states whether it is in effect for this run

3. **Tech stack** - From the project marker files, identify:
   - Frameworks (React, Express, Django, etc.)
   - Databases or data stores
   - Key dependencies that shape how code is written
   - Build tools and test runners

4. **Conventions** - Scan a few representative files to identify:
   - Naming patterns (camelCase, snake_case, kebab-case)
   - File structure conventions (e.g., commands in `.claude/commands/`, skills in `.claude/skills/`)
   - Any patterns called out in CLAUDE.md or rules files

5. **Output a compact summary** - Keep it to 10-15 lines max. Include:
   - Project type and main language
   - Key frameworks and dependencies
   - Critical rules that affect how subagents should behave
   - File structure conventions worth knowing

</procedure>

<rules>

- Keep it lightweight - just enough context for informed decisions
- Do not audit the full codebase - sample a few files
- Do not generate reports or findings - this is context gathering, not review
- If a marker file does not exist, skip it and move on

</rules>

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.