9 Skills That Made My AI Junior Dev 10x Smarter

A few weeks ago, I stumbled on a breakdown of how skills should be structured — nine distinct categories. Reading it, I realized my junior dev agent was using exactly one skill: a bug fix template. One-dimensional output.

Here’s what I learned and applied to my ERP project.

The Problem With a Single Skill

My agent workflow was simple: detect ticket → fix → open PR. It worked, but every ticket got the same treatment regardless of whether the issue was a code bug, configuration drift, data inconsistency, or user error.

Skills aren’t just about telling the agent what to do. They’re about telling it when to do different things.

The Nine Skill Categories — Applied to ERP

Here’s what I built for my project, mapped to that framework:

1. Library and API Reference

My agent often hallucinated internal API endpoints, calling /api/v2/inventory when the real endpoint is /api/v1/stock. The skill documents every internal API with its exact path, auth method, and gotchas.

2. Product Verification

Highest impact. The sanity-check skill runs after every deploy: hit health endpoints, verify DB pool, check job queues, assert API responses. The agent catches regressions before I do.

3. Data Fetching and Analysis

When a user reports “stok laporan error,” the agent needs DB access. The erp-queries skill provides pre-written parameterized queries with read-only credentials — joins, timezone handling, state transitions — without ever writing raw SQL.

4. Business Process Automation

Tickets follow patterns. “User X can’t login” is usually lockout, not a bug. The triage skill classifies issues (bug/config/user-error/feature-request) and routes accordingly, logging every decision for audit.

5. Code Scaffolding

When I add a new ERP module, the new-module skill generates boilerplate — models, migrations, API routes, tests — with conventions pre-wired. Every mistake becomes a skill update.

6. Code Quality and Review

The adversarial-review skill spawns a read-only subagent to critique its own PR. It checks for null handling, user-facing errors, parameterized SQL, and timezone-naive datetimes. Caught three SQL injection vulnerabilities in the first week.

7. CI/CD and Deployment

The deploy skill handles build → migrations → smoke test → switch traffic → monitor → report. The agent reads deploy output, interprets errors, and decides whether to rollback or retry — something a script can’t do.

8. Runbooks

When an alert fires, the oncall runbook: fetches the alert, checks error logs, queries recent deploys, correlates cause, and opens a rollback PR with evidence. This turned my agent from a ticket-fixer into an oncall engineer.

9. Infrastructure Operations

The db-maintenance skill handles vacuum, index health, connection pool adjustments — with guardrails against destructive ops during business hours.

Where the Magic Lives

The highest-signal content across all skills is the Gotchas section. Every gotcha is a mistake the agent made, captured immediately. After two weeks, the agent stopped making those mistakes entirely.

Write Descriptions for the Agent, Not for Humans

The description field is what the agent scans to decide if a skill fits. Wrong: “Inventory library API documentation.” Right: “Use when querying inventory endpoints or debugging stock discrepancies.” The agent now picks the right skill without being told.

Progressive Disclosure

A skill is a folder, not a file. The main skill file points to reference files loaded on demand. My erp-queries skill references separate files for stock, order-tracing, and analytics queries. Before this, the main file was 400 lines of irrelevant query patterns every time the agent needed a simple stock check.

The Result

After two weeks:

• Bugs fixed per week: up from 7 to 12
• False positives (fixing non-bugs): down from 30% to 8%
• PRs requiring rework: down from 40% to 10%

I stopped thinking about whether the agent can handle a task and started thinking about which skill it needs. The setup is the work. The execution is verification.