congmnguyen

Delegate a heavy or mechanical coding task to the configured Codex CLI model via a low-effort Sonnet wrapper, so Codex's large transcript stays OUT of the main orchestrator context. Trigger when the user says "delegate to codex", "hand this to codex", "let codex do X", or when a well-specified implementation/refactor/migration/bulk task is large enough that doing it inline would flood context with tool round-trips. Do NOT use for small tasks you can one-shot.

PyTorch model-building conventions and a neural-net training debug checklist. Use this skill whenever writing or reviewing PyTorch code that defines a model (nn.Linear, nn.Conv2d, BatchNorm) or trains one (training loop, optimizer, LR schedule), and ESPECIALLY when debugging training problems — loss not converging, loss stuck or flat, suspiciously slow learning, "hockey stick" loss curves, or a model that seems to ignore its inputs. Also consult it before starting any from-scratch training run, even if the user doesn't mention debugging.

Tutor the user toward deep understanding of a topic, codebase, paper, artifact, or concept through incremental explanation, active recall, and verification. Use when the user asks to be taught, wants to understand something deeply, says "teach me", "walk me through", "explain until I get it", "quiz me", asks for ELI5/ELI14/intern-level teaching, or needs a learning session instead of a one-off answer. Skip for simple factual answers or implementation tasks where teaching is not requested.

Commit, push, and open a PR

Create native editable draw.io diagrams as .drawio files and optionally export them to PNG, SVG, PDF, JPG, or browser URLs. Use when Codex needs to create, generate, draw, or design a diagram, flowchart, architecture diagram, ER diagram, sequence diagram, class diagram, network diagram, mockup, wireframe, UI sketch, or work with draw.io, drawio, drawoi, .drawio files, or diagram export.

Summarize the current conversation into a handoff file saved to ~/.claude/handoffs/

Automate browser interactions, test web pages and work with Playwright tests.

PyTorch model-building conventions and a neural-net training debug checklist. Use this skill whenever writing or reviewing PyTorch code that defines a model (nn.Linear, nn.Conv2d, BatchNorm) or trains one (training loop, optimizer, LR schedule), and ESPECIALLY when debugging training problems — loss not converging, loss stuck or flat, suspiciously slow learning, "hockey stick" loss curves, or a model that seems to ignore its inputs. Also consult it before starting any from-scratch training run, even if the user doesn't mention debugging.

Commit, push, and open a PR

Tutor the user toward deep understanding of a topic, codebase, paper, artifact, or concept through incremental explanation, active recall, and verification. Use when the user asks to be taught, wants to understand something deeply, says "teach me", "walk me through", "explain until I get it", "quiz me", asks for ELI5/ELI14/intern-level teaching, or needs a learning session instead of a one-off answer. Skip for simple factual answers or implementation tasks where teaching is not requested.

Use when debugging or comparing Text-to-SQL agent eval results in /home/cong/code/text2sql-agent, especially Promptfoo BI20/VNP failures, strict-vs-smoke differences, queryData mismatches, metadata fixes, regression checks, and deciding whether a failure is real or a checker issue.

Use when polishing Vietnamese technical reports, Desktop HTML/Markdown explainers, seminar handovers, monthly work reports, or presentation notes. Focuses on removing internal clutter, rewriting jargon into spoken Vietnamese, preserving meaning, and preparing likely Q&A.