Review a diff as an adversary — assume the code is wrong, work from the diff alone, and try to construct concrete failures. The reviewer-role procedure for multi-agent workflows (one implements, N review), distilled from the Bun-in-Rust port.
Report on your own agent usage — parse Claude Code session transcripts for token burn, and cross-check installed agentstack skills/servers against what was actually used to flag dead weight. Local and read-only.
Find and install agent skills through agentstack — search the library/catalog/registry, pull from any skills repo (owner/repo, git URLs, local dirs), and judge quality by scan verdicts, pins, and provenance instead of install counts. Use when the user asks "how do I X", "is there a skill for X", or wants a capability…
Mine your past agent sessions (Claude Code, Codex) for recurring failures and re-explained procedures, then distill the best candidates into reusable skills that land in the central library via agentstack more lib add. Local and read-only until you approve a draft.
Wire a generate-review-fix multi-agent workflow (one implementer, N adversarial reviewers, a fixer) from agentstack profiles — models per role, skills per role, secrets injected at start — and run it in your executor of choice (sandcastle, Claude Code workflows, or plain Docker). agentstack provisions and governs the…
Turn a product or analytics question into the right PostHog query — trends, funnels, or retention — with the correct events, breakdowns, and date range.
Propose a reviewable multi-agent workflow as a blueprint — pick and name the pattern (map-reduce, pipeline, tournament, loop-until-dry, dag), emit its shape as an agentstack-blueprint JSON block, and WAIT for the user to approve / reject / edit before authoring and running it via agentstack more workflow run.
Pick the right model for each job — route bulk and mechanical work to cheap models, exploration to fast mid models, and reserve the strongest model for judgment, taste, and final review. Escalate when the output misses the bar.
Delegate clear-spec, bulk, or mechanical work — migrations, codemods, large refactors, log digs, and independent code review — to gpt-5.5 through the Codex CLI, with quota-aware fallback to Claude models.
Keep your agentstack central library (/.agentstack/lib) consistent across machines by versioning it as a git repo — commit the index and path-source skill bodies, exclude the content store cache and any resolved secrets.
Operate and troubleshoot AgentStack projects, personal capability libraries, zero-files MCP delivery, toolsets, secrets, trust, policy, machine bootstrap, and legacy MCP cleanup. Use whenever a task changes or diagnoses an AI coding CLI setup managed by AgentStack.