AI Agent Workflow Collection

Use when an agent's performance is uneven and you want systematic improvement. Runs diagnose -> challenge-gen -> challenge-run -> update learnings in a loop. Works with or without a team. Keywords: active-learn, training, adversarial, improve, agent, loop, capability, tuning, solo.

Use when a goal is vague and you need to explore before planning. Runs spikes to discover what work exists, then produces a prioritized backlog. Keywords: explore, discover, spike, plan, epic, backlog, investigate.

Use after /diagnose-agent when you have a struggle profile and need training challenges. Generates edge-case and commit-replay challenges calibrated to weak areas. Keywords: challenge, training, improvement, active-learning, edge-case, replay, agent, weakness.

Use after /challenge-gen when challenges are ready to execute. Dispatches the agent on each challenge and evaluates result, trap detection, and calibration. Keywords: challenge, run, evaluate, active-learning, training, agent, performance, calibration.

Use when the backlog feels unwieldy — after a blossom run, before a sprint, or when you suspect duplicates or stale tasks. Deduplicates, fills gaps, and cleans up dependencies. Keywords: backlog, cleanup, dedup, triage, hygiene, review, organize.

Use when an agent's performance is uneven and you want to understand where it excels or struggles. Produces a struggle profile for /challenge-gen or /active-learn. Keywords: diagnose, profile, agent, weakness, strength, capability, assessment.

Use when deciding between two approaches and you need a structured tradeoff comparison. Produces a dimension-by-dimension table with a recommendation. Keywords: compare, versus, vs, tradeoff, decision, choose, pros cons.

Apply a criterion to items and keep or drop them. The grep of knowledge work — binary keep/drop decision per item. Different from rank (which reorders but keeps all). Keywords: filter, sift, keep, drop, remove, prune, exclude, include, grep.

Use after a human jam session to turn agreements into code. Converts shared artifacts into versioned event schemas, mocks, and validation config. Keywords: formalize, contracts, event schemas, boundaries, multi-human, event storming.

Use when you need deep understanding of a focused area — not broad exploration. Recursively investigates, pruning paths that don't serve the goal. Keywords: explore, research, investigate, understand, deep-dive, discover.

Interactive code review as a panel discussion: an implementer agent defends decisions while a reviewer surfaces findings, with specialist reviewers added based on what the diff touches. Use when you want findings discussed in real time rather than a static report. Keywords: PR review, code review, meeting, panel review, interactive review, review discussion.

Use before a multi-human Event Storming session to prepare independent findings. Each participant's agents explore their domain and produce candidate events, aggregates, and assumptions. Keywords: event storming, domain events, prep, multi-human, exploration, DDD, aggregate, bounded context.

Use when you need to research, analyze, compare, plan, or structure thinking and aren't sure which primitive to reach for. Chains operations (gather, distill, rank, filter, verify, etc.) automatically based on your goal. Prefer this over individual primitives for multi-step work.

Use when a feature crosses system boundaries and integration risk is high. Implements the thinnest end-to-end path first, then widens pass by pass — always working. Keywords: tracer, iterative, incremental, end-to-end, thin slice, vertical slice.