Stages the changes produced in this session, excludes junk and secrets, and generates a conventional commit. Triggered when the user runs /devx-git:ci or asks to commit.
Updates the current branch onto an origin branch via rebase, with a recoverable backup, safe conflict resolution, and recovery steps. Use when the user runs /rebase or asks to rebase/update their branch onto main (or another origin branch).
Audits an agent, tool loop, or inference request across six dimensions (prompt engineering, tools definition, context engineering, cache usage, evaluation, guardrails) and produces a scored gap analysis with severity ratings and concrete remediations. Framework-agnostic; recognizes patterns from Anthropic SDK, OpenAI SDK, Vercel AI SDK, and OpenAI Agents SDK. Triggers on: audit my agent, review tool loop, audit inference request, agentic engineering review, audit prompt, review agent code, check prompt caching, evaluate guardrails.
Rewrites a given prompt or agent definition by applying Anthropic prompt-engineering best practices — clarity and directness, context and motivation, XML structure, role assignment, long-context layout, output format control, communication-style calibration, and parallel-tool-call affordance. Produces a before/after with per-change rationale. Triggers on: optimize this prompt, improve my prompt, rewrite prompt, prompt optimization, refine system prompt, tune agent prompt, polish prompt for production.
Analyzes project architecture and generates ASCII state machine and sequence diagrams. Use for explaining a project, analyzing architecture, tracing an entrypoint, showing a state machine, or generating a sequence diagram.
Audits a project's agentic harness (CLAUDE.md, AGENTS.md, .claude/rules, .claude/settings.json) and refactors it into a layered, path-scoped configuration that loads only what is relevant. Rewrites CLAUDE.md and AGENTS.md as a thin critical-rules layer, segments domain rules with `paths:` frontmatter, and adds deny-first permission guardrails. Use for: fixing the harness, auditing claude.md, refactoring agentic configuration, improving agent rules, segmenting rules by path, reviewing settings.json, harness review, or when claude.md is too long.
Evaluate and improve a plugin's skills by applying Anthropic's skill authoring best practices. Use to fix, improve, or review skills for a specific plugin or agent team.
This skill should be used when the user asks to "security review this PR", "review this diff for vulnerabilities", "appsec review these changes", "audit this branch for auth, injection, or data exposure bugs", "check this code for exploitable security issues", or requests a focused application security review of pending changes.