Creating Skills

Prerequisites

Before creating a skill:

Understand the issue you want to detect (specific bug patterns, architectural problems, security flaws)
Have example code that demonstrates the issue
Know what context the skill needs to make accurate judgments

Quick Start

Create a skill in three steps:

Create the skill directory

mkdir -p .agents/skills/my-skill

Write SKILL.md

touch .agents/skills/my-skill/SKILL.md

Add to warden.toml

[[skills]]
name = "my-skill"

[[skills.triggers]]
type = "pull_request"
actions = ["opened", "synchronize"]

Skill Template

Start with this template:

.agents/skills/my-skill/SKILL.md

---
name: my-skill
description: "Brief one-line description for discovery"
allowed-tools: Read Grep Glob
---

You are an expert [domain] analyzer.

## Scope

You receive code chunks from diffs. Analyze each chunk against the checks below.

## Checks

### Check 1: [Specific Pattern]

**What to look for:**
- Concrete code pattern 1
- Concrete code pattern 2

**Red flags:**
- Accessing X without checking Y
- Using Z when condition W is not met

**Safe patterns:**
- Using helper function `foo()` which handles edge cases
- Explicit check before operation

**Not a bug:**
- False positive clarifications

---

### Check 2: [Another Pattern]

[Same structure as Check 1]

## Report Format

For each finding:
- File path and line number
- Which check it matches
- One sentence: what is wrong
- Why it matters (trigger condition, failure mode)
- Suggested fix (only if obvious)

### Zero findings

If no checks fire, report nothing. Silence means the code is clean.

## Severity Levels

- **high**: Will cause incorrect behavior, data loss, or crash in normal usage
- **medium**: Incorrect behavior requiring specific conditions
- **low**: Minor issues or style concerns

Real Example: SQL Injection Detection

Let’s build a skill that detects SQL injection vulnerabilities.

Define the pattern

We want to catch:

String concatenation in SQL queries
User input passed directly to query functions
Missing parameterization

Write the skill

.agents/skills/check-sql-injection/SKILL.md

---
name: check-sql-injection
description: "Detect SQL injection vulnerabilities from improper query construction"
allowed-tools: Read Grep Glob
---

You are a security-focused code reviewer specialized in SQL injection detection.

## Scope

Analyze code for SQL injection vulnerabilities. Focus on database query construction.

## Check: SQL Injection via String Concatenation

**Red flags:**
- String concatenation or template literals building SQL with user input: `query = "SELECT * FROM users WHERE id = " + userId`
- `.format()` or `f-strings` interpolating variables into SQL
- User-controlled values in `WHERE`, `ORDER BY`, `LIMIT` clauses without parameterization
- Raw SQL execution functions receiving concatenated strings

**Safe patterns:**
- Parameterized queries: `db.query('SELECT * FROM users WHERE id = ?', [userId])`
- ORM methods with bound parameters: `User.findOne({ id: userId })`
- Query builders with parameter binding: `knex('users').where('id', userId)`
- Hardcoded SQL strings without user input

**Not a bug:**
- Concatenation of hardcoded SQL fragments (table names, column names from constants)
- Logging or error messages containing SQL (not executed)
- SQL in comments or documentation

## Report Format

For each vulnerability:
- File path and line number
- The vulnerable code line
- What user input flows into the query
- Why parameterization is needed
- How to fix (show parameterized version)

## Severity

- **critical**: User input directly concatenated into executed SQL
- **high**: Indirect flow of user input into SQL without validation
- **medium**: User input with basic sanitization (still vulnerable)

Test the skill

Create test files with vulnerable patterns:

test.ts

// Vulnerable - should trigger
const userId = req.params.id;
const query = `SELECT * FROM users WHERE id = ${userId}`;
db.query(query);

// Safe - should not trigger
const result = db.query('SELECT * FROM users WHERE id = ?', [userId]);

Run Warden locally:

warden run --skill check-sql-injection

Iterate and refine

Based on results:

Add more patterns to catch variants
Refine “safe patterns” to reduce false positives
Adjust severity based on real impact

Writing Effective Checks

Be Specific

Good vs Bad specificity

Too vague:

Check for improper error handling

Too specific:

Check if error at line 42 of auth.ts is caught

Just right:

Check if async functions catch promise rejections.
Red flags:
- Async function without try/catch
- Promise chain without .catch()
- await without surrounding try/catch

Provide Examples

Show what triggers the check and what doesn’t:

**Red flags:**
- `response.content[0]` without checking array length
- `msg.usage.input_tokens` without null check

**Safe patterns:**
- `response.content[0] ?? defaultBlock`
- Checking `response.content.length > 0` first

Explain Impact

Help the user understand why it matters:

**Trigger**: When the API returns an empty content array, accessing `[0]` throws TypeError, crashing the application.

Calibrate Confidence

Set clear thresholds for reporting:

## Confidence Calibration

| Level | Criteria | Action |
|-------|----------|--------|
| HIGH | Pattern traced to specific code, confirmed triggerable | Report |
| MEDIUM | Pattern present, but context may mitigate | Read more context, then decide |
| LOW | Vague resemblance | Do NOT report |

When in doubt, read more files. Never guess.

Advanced Patterns

Multi-Step Analysis

For complex checks, guide the agent through steps:

## Step 1: Classify the Code

Identify which architectural zone(s) the code touches:
- SDK layer (`src/sdk/`)
- Config layer (`src/config/`)
- Output layer (`src/output/`)

Only run checks relevant to the touched zones.

## Step 2: Run Checks

### Check 1: SDK Response Handling
(Only if SDK layer is touched)
...

## Step 3: Report

For each finding:
...

Historical Context

Reference past bugs to sharpen detection:

### Check 2: Config Threading

**Historical commits:** 8+

Config flows through a 3-level merge chain. Breaking this chain causes silent feature failure.

**Red flags:**
- New config field not threaded through `resolveSkillConfigs()`
- Using `||` instead of `??` when 0/false are valid values
...

Zone-Based Scoping

Skip irrelevant checks based on file paths:

## Step 1: Classify the Code

Before running checks, identify which zones the code touches:
- **SDK layer** (`src/sdk/`): Response parsing, IPC
- **CLI layer** (`src/cli/`): Task orchestration, rendering

Only run checks relevant to the touched zones.

Testing Skills

Local Testing

Test against specific files:

# Run on staged changes
warden run --skill my-skill

# Run on specific diff
git diff main | warden run --skill my-skill --stdin

# Run on specific files
warden run --skill my-skill --files src/problematic.ts

Create Fixtures

Build a test suite for your skill:

.agents/skills/my-skill/
├── SKILL.md
└── fixtures/
    ├── vulnerable.ts     # Should trigger finding
    ├── safe.ts          # Should not trigger
    └── edge-case.ts     # Tricky case

Iterate on Real Code

Run the skill on your actual codebase:

# Full repo scan
warden run --skill my-skill --schedule

# Check recent PRs
gh pr list --limit 5 | while read pr; do
  gh pr diff $pr | warden run --skill my-skill --stdin
done

Common Pitfalls

Avoid these anti-patterns when writing skills:

Too generic: “Check for bugs” → Specify exactly what constitutes a bug
Style preferences: Skills should find correctness issues, not enforce formatting
False positive factories: Add “safe patterns” and “not a bug” sections
No severity guidance: Define what makes an issue high vs medium vs low
Implicit context: Don’t assume the agent knows your codebase—explain patterns

Prompt Engineering Tips

Role Framing

Set the right persona:

✅ "You are an expert bug hunter who knows Warden's architecture intimately."
❌ "You are a helpful assistant."

Structured Output

Guide the report format:

For each finding:
- File path and line number
- Which check (1-9) it matches
- One sentence: what is wrong
- Trigger: specific condition that causes failure
- Suggested fix (only if clear)

### Zero findings

If no checks fire, report nothing. Do not invent findings.

Calibrated Confidence

Prevent noise:

Do NOT report low severity findings. If confidence is that low, don't report it.

Get Started

Core Concepts

CLI Usage

GitHub Action

Configuration

Skills

Advanced

Creating Skills

Prerequisites

Quick Start

Skill Template

Real Example: SQL Injection Detection

Writing Effective Checks

Be Specific

Provide Examples

Explain Impact

Calibrate Confidence

Advanced Patterns

Multi-Step Analysis

Historical Context

Zone-Based Scoping

Testing Skills

Local Testing

Create Fixtures

Iterate on Real Code

Common Pitfalls

Prompt Engineering Tips

Role Framing

Structured Output

Calibrated Confidence

Next Steps

Skill Structure

Builtin Skills

Build docs developers (and LLMs) love

Get Started

Core Concepts

CLI Usage

GitHub Action

Configuration

Skills

Advanced

​Prerequisites

​Quick Start

​Skill Template

​Real Example: SQL Injection Detection

​Writing Effective Checks

​Be Specific

​Provide Examples

​Explain Impact

​Calibrate Confidence

​Advanced Patterns

​Multi-Step Analysis

​Historical Context

​Zone-Based Scoping

​Testing Skills

​Local Testing

​Create Fixtures

​Iterate on Real Code

​Common Pitfalls

​Prompt Engineering Tips

​Role Framing

​Structured Output

​Calibrated Confidence

​Next Steps

Skill Structure

Builtin Skills

Build docs developers (and LLMs) love

Prerequisites

Quick Start

Skill Template

Real Example: SQL Injection Detection

Writing Effective Checks

Be Specific

Provide Examples

Explain Impact

Calibrate Confidence

Advanced Patterns

Multi-Step Analysis

Historical Context

Zone-Based Scoping

Testing Skills

Local Testing

Create Fixtures

Iterate on Real Code

Common Pitfalls

Prompt Engineering Tips

Role Framing

Structured Output

Calibrated Confidence

Next Steps