AdamKrysztopa

Use when setting up or repairing the Python env for analysis — detects uv/venv/conda/poetry/pipenv, installs core packages

Use when building or improving a predictive model — AIDE-style solution tree with leakage discipline and empirical leaderboard

Use when a data question is fuzzy or high-stakes — clarifies scope and writes analysis-spec.md before running a solver

Use when starting fresh with data and unsure which skills to use — peeks at data, asks targeted questions, assembles a crew plan

"Use when onboarding a dataset or asking 'what\'s in this data?' — per-column quality report with join-compatibility checks; flags possible PII/sensitive columns"

Use when browsing, pruning, or seeding a run from past analyses stored across sessions

Use when a single hard task keeps failing greedily — tree-searches alternative solution paths via MCTS mode

Use when a number must be right — runs N diverse data scientists in parallel and reconciles into consensus + minority report

Drive ONE SciAgent backlog story from docs/reference/core.md to verified, gate-green, committed code — goal-directed end-to-end. The controller (Opus) takes the next story off the execution order, branches, gathers the contract (story + CLAUDE.md invariants + Context7 for any SDK), then executes it subagent-driven — implementation ALWAYS on cheaper models (Sonnet default, Haiku for fully-pinned mechanical tasks), Opus NEVER writes code. When ruff + pyright + pytest and the story's acceptance criteria are all green (verified by the verification-gate agent), it commits and reports. Use when the user runs /deliver-story or asks to build/ship/deliver a story (e.g. "AGT-12", "the gap-finder story", "the next story") from docs/reference/core.md.

Regenerate the typed TypeScript SciAgent API client(s) from the backend's live OpenAPI schema, so the Zotero add-on and the mechai.pl web app stay in lockstep with the FastAPI backend instead of drifting against hand-written types. Use when src/agt/api/** or src/agt/models.py changed, when api-contract-guardian reports drift, or when bootstrapping a new client (the mechai.pl web app). Run /gen-api-client.

Background policy for which Claude model tier to use when delegating SciAgent work to subagents — plan/architect on Opus, implement on Sonnet, fully-pinned mechanical tasks and gate-running on Haiku. Applies whenever Claude spawns subagents for SciAgent work, even outside the full deliver-story pipeline, so cost/quality tiering happens automatically without the user restating it.

Translate docs/reference/core.md epics into delivery-ready tickets, implementation plans, and validation checklists.

Apply docs/reference/settings.md to bootstrap local development and quality tooling in a reproducible way.

Build and validate Zotero plugin features based on docs/reference/zotero.md milestones and constraints.

Use when designing or assessing an LLM-agent system — 'how should I build this agent', 'do I need multiple agents or one', 'should this be a workflow or an agent', 'how do I add memory/tools/human approval', 'is my ReAct loop / multi-agent setup right', or any agent-architecture code review. Branches automatically: greenfield → a layered design interview (autonomy → reasoning loop → topology → memory → reliability → governance → integration); existing agent code → a review against the catalog's seven recurring defects. Pushes toward the least autonomy that works. Reach for this whenever agents, tools, multi-agent, RAG-agents, orchestration, or LLM control flow come up, even if no pattern is named.

Use when choosing or assessing software architecture — picking a structure/topology for a new system, or reviewing whether an existing codebase's architecture fits. Triggers on questions like 'how should I structure this service', 'should this be microservices or a monolith', 'is our layering right', 'we're adding events/CQRS/a queue — does it fit', or any greenfield architecture decision or architecture-focused code review. Branches automatically: greenfield → a short selection interview that composes a stack; existing code → a review against the pattern catalog with targeted moves. Reach for this even when the user hasn't named a specific pattern.

Use when choosing or assessing object-level design patterns (GoF and Python-idiomatic) — 'which pattern fits here', 'how should I structure this class', 'is this the right use of a factory/strategy/observer', 'how do I make these algorithms swappable', or any design-pattern code review. Branches automatically: greenfield → a short interview that returns one recommended pattern (with its Pythonic form); existing code → a review that maps smells to the patterns that fix them. Strongly prefers the language feature over the ceremony — reach for this whenever pattern/class-design questions come up, even if no pattern is named, and especially in Python.

Dispatch the orchestrator subagent for any task in the dependence-forecastability (ami) project. Use for every code change, bug fix, test run, review, planning, or release — even one-liners. Invoked as /ami-orchestrator <task>. Pre-loads pytest -n auto (parallel, never piped), active release plan, and all stage-gate constraints so the orchestrator gets full project context from the first message.