Eight production-grade Agent Skills that stop the sycophancy, the slop, and the footguns -- so the agent's output is actually safe to ship. Markdown procedures you read in a minute and use for years.
Get the pack See what's insideA coding agent will happily agree with your bad idea, write code that passes the tests but is not safe to ship, run a destructive shell command on a long unattended run, and loop forever declaring false victory. These eight skills are the guardrails.
Eight skills, each a single SKILL.md your agent loads on demand. No install step, no dependencies, nothing to run -- they encode how to get reliable results.
Defeat sycophancy -- get a real judgment on an idea, plan, or the agent's own work instead of the model agreeing with you.
Review AI-generated code like a senior developer: a 9-step ordered review + pre-merge checklist that catches bugs that pass the tests but are not safe to ship.
Build an unattended agent loop that prompts itself, checks its own work, and stops or escalates -- with the non-negotiable escape hatch.
Choose the smallest sufficient context per step instead of dumping the whole window, so the agent stays sharp and cheap.
Write a CLAUDE.md the agent actually follows -- a behavioral contract of imperative rules, each closing a specific failure mode.
Make a multi-agent system converge on a finished result (Researcher, Builder, Judge, Manager; builder is not the judge; bounded revision).
Lock down the most dangerous tool an agent has -- the shell -- so a long autonomous run cannot do something irreversible.
Turn any repeating workflow into a durable Agent Skill in the real Anthropic format, and refine it so it compounds.
Developers and teams already using Claude Code, Codex, or another coding agent day to day, who are past "it writes boilerplate" and want the agent's work to be reliable enough to trust. If you have never opened a coding agent, start there first -- this pack assumes you already run one.
1. Buy and download the pack. 2. Drop a skill folder into ~/.claude/skills/
(all projects) or your repo's .claude/skills/ (one project). 3. Invoke it by name
or let the agent auto-load it when your request matches. That's it -- they're plain markdown.
You need to already use a coding agent (Claude Code or similar). The skills assume you run one; they make its output reliable.
No. Each skill is a single markdown file (a procedure the agent reads). You copy it into a skills folder -- no dependencies, nothing to run.
It's written for Claude Code, but the Agent Skill format is plain markdown, so it works with any harness that reads skills. The techniques are largely model-agnostic.
The license is single-seat -- use it yourself and on your own projects. Reselling or redistributing the files is not permitted. Ask about a team license.
Yes -- buyers get free updates to this pack.