human-training

Use when driving Google's Antigravity CLI (`agy`) non-interactively / headlessly — `agy -p`/`--print`, running it from a script or another agent, or automating a Gemini-backed agent call from the terminal. Triggers include "run agy from the command line", "antigravity headless / non-interactive", "agy -p", "is there an antigravity exec/print mode", "automate antigravity", "agy prints nothing / empty output / no stdout", "pick the agy model", and using agy/Gemini as a cross-vendor consultant or reviewer over files. Do NOT use for the interactive `agy` TUI, the Antigravity IDE (Electron app), Antigravity 2.0 desktop, the Antigravity SDK (Python), or the hosted Gemini API (that's the gemini-api skill).

Work out what a specific change breaks somewhere else, beyond the diff, and prove the one fact its safety depends on by running code rather than writing it up. Triggers on /blast-radius, "blast radius of X", "what could this break", "is this safe to change", "what else touches this", or reviewing a small diff you don't trust. Do NOT use for a wide audit of existing code with no particular change in view — that's robustness-audit.

Use when driving OpenAI's Codex CLI non-interactively / headlessly — `codex exec`, running Codex from a script or another agent, capturing its output to a file, or automating a Codex call. Triggers include "run codex from the command line", "codex headless / non-interactive", "codex exec", "is there a codex -p", "automate codex", "codex hangs / stuck with no output", "capture codex output", "codex in CI". Do NOT use for the interactive Codex TUI, the Codex desktop GUI app, or for the OpenAI cloud/Responses API or non-Codex tools.

Iterate an artifact whose quality has no test — only a perceiver — against fresh, blind critic agents. On invocation, designs a harness for THIS artifact (rubric, gate, budget in rounds, walls for what the critic cannot see) and proposes it for approval before any round is spent; then runs a free inner loop and rationed critic rounds, logs a per-line vector, and treats a documented stop as a deliverable. Model- and artifact-agnostic. Triggers include "does this read as X", "would a stranger get this", "am I too close to this", "fresh eyes", "critic round", "review loop", "iterate with a reviewer", "is this recognisable / persuasive / clear / obvious what it's for". Explicit invocation via /critic-loop. Not for qualities that have a test — use the test.

Use when writing, reviewing, debugging, or migrating Gemini API or hosted-Gemma code (Gemini 3.x, Gemma 4). Triggers on imports of `google.genai` / `@google/genai`, model strings like `gemini-3-flash`, `gemini-3.5-flash`, `gemini-3.1-flash-lite`, `gemma-4-31b-it`, `gemma-4-26b-a4b-it`, calls to `generate_content` / `generateContent`, config of `GenerateContentConfig` / `ThinkingConfig`, and requests to wire up Gemini or Gemma for a task. Use proactively before writing Gemini/Gemma code, so defaults reflect current 3.x conventions rather than older training-data ones. Do NOT use for general LLM-provider comparisons, OpenAI/Anthropic specifics, locally-hosted Gemma open weights (Ollama / HuggingFace / vLLM — this covers the hosted API path), or non-Gemini Google Cloud.

Adversarial pre-build alignment. Interrogates you one question at a time — recommended answer on each — walking every branch of the decision tree until "done" is crisp and bounded, then sharpens fuzzy domain terms into CONTEXT.md inline. The fix for a fuzzy plan, creeping scope, or unclear done-criteria. Explicit invocation via /grill before starting a change; use when you want to be grilled, stress-tested, or pinned down before writing code.

Use when the user requests a handover, the context window is getting full or laggy, at a natural pause point (end of sprint or milestone), or when stuck and a fresh perspective is needed. Creates a handover capturing the ephemeral conversation delta not already in the project files. Works with any AI agent.

Use when the user wants to CREATE or EDIT a raster/photographic image — 'generate an image', 'make a photo/logo/hero/mockup/sprite/icon-art', 'edit this picture', 'change/add/remove/replace X in this image', img2img, style or background swaps, identity-preserving edits, upscale-style redo. Claude CAN do this (like ChatGPT/DALL·E) by driving the Codex CLI's built-in image tool — no API key, runs on a ChatGPT subscription. Do NOT use for code-native or vector visuals (SVG icons, charts/dataviz, CSS/HTML/canvas UI, algorithmic/generative art) — those have their own skills; and do NOT use for the OpenAI image API directly (this is the subscription/CLI path).

Autonomy-biased operating mode for converting otherwise-wasted token quota (near 5-hour or weekly window resets) into useful work. Runs the goal you state — or, if you state none, delegates to project-checkup or robustness-audit — with a relaxed confidence bar, logging decisions to a trail instead of pausing for moderate-confidence calls. Can run unattended for a long single turn, using subagents to stay context-lean. Explicit invocation via /human-training:leroy-jenkins (Claude Code) or $human-training:leroy-jenkins (Codex).

Use when actively implementing features in Sprint 1-N: writing code, writing tests right after, updating the DEVLOG with decisions and rationale, checking the confidence bar before acting, orienting at the start of a session, or handling a mid-sprint context reset. The core test-code-document development loop.

Use when setting up a new project, after the Sprint 0 spec is complete, or when an AI agent entry point is missing. Creates onboarding.md, the universal orientation doc that works across any environment (Cursor, VSCode, Claude Code, web) - the office tour, not the employee handbook.

Split a large PDF into smaller PDFs along an editable, structure-aware plan, preserving and rebasing bookmarks in each chunk. Use when a PDF is too large to process and must be broken into chapters/sections (for reading, LLM ingestion, or RAG), when splitting a book/manual/standard by its table of contents, or when a bookmark-less PDF needs its structure inferred or simple page-based splitting. Scales to 6000+ page documents; when there is no TOC it falls back to heading inference, fixed-page splitting, or a hand-edited page list.

State-aware project health check that composes workflow-orientation, robustness-audit, a "next-move" inventory, and a friction audit. Designed to minimize re-entry friction on dormant or abandoned hobby projects. Explicit invocation via /project-checkup.

Use when starting a new project, brainstorming an idea, evaluating whether something is worth building, or turning a concept into a technical spec. Covers Sprint 0 end to end: ideation (challenge ideas, red-team, force scope) and spec writing (concrete specs for coding agents). Enter at brainstorming, enter at speccing, or flow through both.

Find latent runtime bugs in existing code by close reading, without executing it — crashes, swallowed errors, edge cases, resource leaks, races, and auth gaps that pass tests but fail in production. Triggers include robustness audit, hardening pass, defect-class or FMEA-style review, "what could go wrong", "what's likely broken", "this feels brittle", silent or mysterious failures, bugs that pass CI, features that work in tests but not in production, auditing before a release, and assessing a codebase after time away.

Keep a reviewable decision trail for work a human reviews after the fact — unattended runs, multi-phase efforts, anything they stepped away from. One append-only TSV, a row per decision, written as decisions happen rather than assembled at the end. Owns the trail format that other skills compose with. Triggers on /show-me-your-work, "keep a decision trail", "log your decisions so I can review", or an autonomous run that needs an auditable record.

Explicit fresh-session opener for Claude Code (/human-training:start) or Codex ($human-training:start). Reads the project's temporal docs and in-repo memory index, sweeps ticked TASKS Active items to Done, checks unpushed commits and local Claude memory that may be stranded, then proposes the next move. Claude Code also checks marketplace freshness. Docs-only; the TASKS sweep is its one write. Use only when the user selects this skill at the top of a fresh session.

Use when managing a project's task list / todos — adding, reviewing, or completing tasks, or parking a tangent or "I should also…" idea so it doesn't derail current work. Maintains a per-project TASKS.md (Active / Someday / Done). Triggers on /tasks, "add a task", "what's on my list", "park that for later", "what was I going to do next". Offers to capture commitments and tangents mid-work, always confirming first. Do NOT use for recording decisions or rationale (that's the DEVLOG) or for workplace contact/jargon memory.

Use when entering a project to align it with the sprint-based workflow: empty project (scaffold rails), existing project with no workflow (propose onboarding), partial setup (gap report), canonical-healthy (drift check), or mature project with its own conventions (bridge mode via onboarding.md). Also owns project memory placement: migrating memory out of a harness's global path (~/.claude/projects/...) and into the repo, for private repos only. Always audits read-only first, discusses, then acts non-destructively.