AojiLi

Set up or update a repository with native Codex AGENTS.md guidance plus optional on-demand CONTEXT.md and repo-local skills. Use when the user wants to initialize a repo for long-term Codex work, create or repair Codex settings, adapt the Codex Project Settings Framework to an existing or empty project, add reference-backed project guidance, reduce project-instruction context, or ensure Codex has explicit repository coverage before writing durable settings.

Investigate a question against high-trust primary sources and capture the findings as a Markdown file in the repo. Use when the user wants a topic researched, docs or API facts gathered, or reading legwork delegated to a background agent.

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".

Review non-trivial software engineering decisions with explicit repository coverage before giving advice. Use when the user asks for architecture, refactoring, implementation strategy, module design, technical tradeoff, risk review, or engineering plan recommendations and the answer depends on understanding an existing repository or turning a software idea into an executable engineering path.

Create fast, desktop-first, Apple-style HTML visual explainers from user-provided concepts, formulas, documents, or software repositories. Use when the user wants a polished standalone HTML explanation, interactive teaching page, architecture walkthrough, codebase map, process or data-flow visualization, or explicitly invokes apple-html-explainer. Default to one self-contained index.html, strong Apple-inspired typography and composition, focused diagrams, restrained interaction, and one quick desktop browser check. Do not default to mobile verification, full web-app states, multi-agent work, broad research, or gesture-physics engineering.

Clarify vague ideas through dialogue, ask for explicit approval before multi-agent research fan-out, automatically challenge assumptions with skeptic loops after synthesis, and produce feasibility, value, and implementation conclusions. Use when the user has an idea, proposal, product concept, research direction, project plan, or architecture thought and wants to know whether it is reasonable, valuable, differentiated, or how to execute it.

Analyze any user decision, choice, tradeoff, option selection, or whether-to-do-something question. Use when the user asks for decision help, wants a recommendation, or invokes decision-advisor. The workflow clarifies the decision, candidate options, background, purpose, core question, jurisdiction, constraints, and decision weights; runs research-backed analysis through economics, strategy, psychology, risk management, legal/regulatory, and technical feasibility lenses; applies critique-revision loops; and returns a clear recommendation, alternatives, action plan, and review signals.

Edit, polish, correct, complete, or insert user-written text while preserving the user's voice, meaning, tone, warmth, and personal style. Use when the user asks to improve grammar, wording, clarity, accuracy, flow, or missing content in a paragraph, selected passage, Markdown/text document, README, DOCX, Google Doc, or other prose document, especially when they want the writing to stay recognizably theirs rather than be heavily rewritten.

Use when training, evaluating, tuning, or analyzing the v54/v55 Level3 low-level PPO reference tracker as a trajectory-following controller, before planner integration. Use for tracker qualification curricula such as hover, point hold, point reach, straight-line tracking, L-shaped tracking, curved trajectory tracking, braking-to-point, heading tracking, multi-point reference following, optional gate-aperture diagnostics, and deciding whether the tracker is ready for a planner-driven Level3 run. Use this instead of level3-ppo-loop when the question is about proving the bottom PPO can accurately, smoothly, and stably follow reference points/trajectories rather than directly optimizing Level3 gate-pass success.

Use when tuning, evaluating, or diagnosing the deterministic GeometricSlowGatePlanner gate-crossing strategy for Level3 on unchanged config/level3.toml, especially v56_geometric_gate_crossing_tuning_loop work involving align stabilization, gate-local speed metrics, cross speed, near-plane backout, recover-after-environment-target-gate-switch logic, per-step trace metrics, or fixed seeds 101-120 planner smoke. Use this instead of level3-ppo-loop when the task says not to train PPO and only to adjust geometric planner/reference generation.

Use when implementing, evaluating, tuning, or analyzing the Level3 MPPI oracle/teacher loop for drone racing on unchanged config/level3.toml, including MPPI controller work, non-PPO controller evaluation, hard-eval reports, and teacher-data gating for later PPO imitation. Use instead of level3-ppo-loop when the task is about MPPI, sampling-based control, oracle controllers, teacher trajectories, BC/DAgger data from MPPI, or checking whether MPPI can pass Level3 before PPO training.

Use when training, evaluating, or tuning the Level3 PPO drone-racing loop for final evaluation on config/level3.toml with the target of at most 7.0s mean successful time and at least 60 percent success.