skill

claude-opus-agents

5 production-grade Claude Code subagents tuned for Opus 4.8: cost-sentinel, threat-modeler, incident-commander, migration-pilot, proof-gate. Validated in CI.

aharwelik/claude-opus-agentscommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/aharwelik/claude-opus-agents itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 23, 2026

claude-opus-agents

Five production-grade Claude Code subagents, each built for a job people actually struggle to get an AI to do reliably — and each tuned for Claude Opus 4.8. Drop them into .claude/agents/ and invoke them by name.

These aren't "you are a helpful assistant" wrappers. Every agent has a sharp scope, hard rules, and an evidence-first output contract.

The agents

Agent	Solves	Hard rule
cost-sentinel	LLM token waste & runaway-spend patterns in code/prompts/agent loops	Ranks by expected cost impact; never claims "cost-safe"
threat-modeler	Generic, ungrounded security reviews	STRIDE mapped to real code; verifies a control exists before crediting it
incident-commander	Chaotic, guess-driven incident response	Read-only — proposes state-changing actions, never executes them
migration-pilot	Big-bang upgrades that break everything	Many small reversible steps; stops on the first red
proof-gate	"It should work" hand-waving	Refuses to return PASS on unproven criteria

Why Opus 4.8

Each of these leans on traits that a top-tier model brings:

Long context lets threat-modeler and incident-commander reason over a whole subsystem (many files, full log windows) instead of fragments.
Strong instruction-following is what makes the strict output contracts (verdict tables, read-only discipline, "never PASS unproven") hold up under pressure.
Better judgment under ambiguity is exactly where proof-gate and cost-sentinel earn their keep — separating proven from plausible.

model: opus is set in each agent's frontmatter so they run on the strongest tier by default.

Install

# Project-scoped
mkdir -p .claude/agents && cp agents/*.md .claude/agents/

# Or globally for every project
mkdir -p ~/.claude/agents && cp agents/*.md ~/.claude/agents/

Then in Claude Code, invoke by name (e.g. "use threat-modeler on the new payments endpoint") or let Claude delegate automatically based on each agent's description.

Validation

python3 tests/validate.py     # 5 passed, 0 failed

CI validates every agent on each push: well-formed frontmatter, required keys (name, description), name-matches-filename, unique names, and a non-trivial system prompt.

Author

Anthony Harwelik — security & AI engineering aharwelik@gmail.com · github.com/aharwelik