agent ops discipline

Long-running AI coding agents often drift into inefficiency, creating redundant tools, letting documentation rot, or burning through token budgets on trivial tasks. This skill enforces a rigorous operational discipline for agents working on complex, multi-week projects. It provides a framework for resource management, task hygiene, and cross-model routing to ensure your agent acts like a senior engineer rather than a runaway script.

What it does

• Redundancy prevention mandates checking existing package.json scripts, Makefiles, and tool directories before any new code is proposed.
• Stale-data protection prohibits hardcoding values in documentation, requiring generated numbers or tests that fail when data drifts.
• Token cost optimization routes work to the cheapest capable model based on the specific deliverable rather than file complexity.
• Tracker integrity enforces aggressive ticket management, ensuring task statuses, assignments, and descriptions reflect the actual state of play.
• Environment cleanup ensures all background processes, dev servers, and temporary git branches are purged immediately upon task completion.

How it works

1. Pre-flight check: Before building a new feature or tool, the agent scans the project index and synonyms to verify the capability doesn't already exist.
2. Model routing: The agent evaluates whether a task is mechanical, implementation-heavy, or architectural to select the appropriate model tier.
3. Execution & brief: Subagents are dispatched with tight, scoped briefs that include explicit "done when" conditions and cleanup rules.
4. Finalization: Once verified, the agent merges code, deletes the local and remote branches, and closes the associated tracker ticket.

Frameworks & tools

This skill is platform-agnostic but designed to interface with Git, task trackers like GitHub Issues, Linear, or Jira, and various LLM tiers (small, mid, and top-tier reasoning models).

Why this beats prompting it yourself

Generic prompts often fail to address the "hidden costs" of agentic workflows, such as orphan background processes or context window bloat from stale instruction files. This skill codifies specific operational failures learned from months of real-world production use, preventing expensive re-work and human-interruption cycles.

Use cases

• Managing a complex refactor across multiple PRs without leaving stale branches.
• Keeping a project README updated with real build metrics and test counts.
• Orchestrating a team of subagents to handle routine mechanical edits at low cost.
• Maintaining a clean Linear or GitHub project board during high-velocity development.

Known limitations

Does not provide automated integration with specific tracker APIs; the agent must have its own tools to interact with Jira or Linear. It cannot prevent a human from manually overriding its cleanup rules.