javi-ai

Interactive tutorial-style lessons for AI agent skills. Instead of passively reading SKILL.md, invoke /powerup <skill> for a guided walkthrough with examples, exercises, and knowledge checks. Trigger: When onboarding to a new skill, teaching a concept interactively, or when a user says "powerup", "teach me", "walkthrough", or "tutorial".

Self-installing markdown recipes for AI agent skills. A pattern where skills include an INSTALL.md that an agent reads and executes to set up dependencies, MCP servers, or config. Trigger: When creating new skills with setup requirements, packaging skills for distribution, or onboarding users to skills that need external dependencies.

Four quality gate checklists for UI polish: baseline visual review, accessibility (keyboard/focus/ARIA), metadata (titles/meta/social cards), and motion performance. Run these gates before shipping any user-facing change. Trigger: When reviewing UI changes, preparing for release, auditing accessibility, or optimizing web performance.

Browser-based UI verification using Playwright. Covers Page Object Model, selector best practices (data-testid, role-based), visual regression, network interception, MCP integration, and CI pipeline configuration. Trigger: When writing E2E tests, verifying UI behavior, setting up Playwright, or debugging flaky browser tests.

Validate that implementation matches specs, design, and tasks. Trigger: When the orchestrator launches you to verify a completed (or partially completed) change.

Definition-of-Done checklist enforcer for SDDs. Detects the change "shape" (new column, new admin page, new CHECK constraint, frontend onChange form) and runs the matching shape-specific checklist before allowing sdd-archive to mark a change COMPLETE. Captures hard-earned lessons from real shipped bugs to prevent the same class of failure on future SDDs. Trigger: invoked by sdd-verify before persisting verify-report, or manually via /sdd-dod {change-name}.

Process markdown notes to auto-insert [[wikilinks]] for key concepts and create stub concept notes. Trigger: When linking notes, extracting concepts, or user says "auto wikilink", "link concepts", "extract entities".

Detect contradictions when synthesizing or consolidating notes. Flags factual disagreements, date conflicts, and outdated info with [!contradiction] callouts. Trigger: When consolidating notes, merging knowledge, or user says "find contradictions", "check conflicts", "contradiction scan".

Autonomous nightly knowledge consolidation — entity deduplication, citation repair, orphan cleanup, and memory compression. Runs as a scheduled task via Claude Code's /schedule or cron, simulating how the brain consolidates memories during sleep. Trigger: When user says "dream cycle", "nightly consolidation", "knowledge cleanup", "run maintenance", or configures a scheduled consolidation task.

Convert markdown document content (notes, docs, wikis) into Mermaid diagrams. Unlike code-to-diagram tools, this generates diagrams from PROSE — summarizing document structure, extracting concept relationships, and visualizing timelines. Trigger: When user says "diagram this doc", "mermaid from notes", "visualize document", "mindmap from markdown", "flowchart from notes", or wants to turn prose into diagrams.

Detect knowledge base health issues in an Obsidian vault — orphan notes, dead wikilinks, missing backlinks, duplicate entities, uncategorized notes, and stale content. Trigger: When auditing vault health, cleaning up notes, or user says "wiki lint", "vault health", "orphan notes".

DSPy-inspired declarative output constraints for SDD sub-agents. When outputs violate assertions, agents self-refine automatically. Adds runtime guardrails without manual review bottlenecks. Trigger: When adding output constraints to SDD agents, enforcing artifact quality, or building self-correction loops.

DSPy-inspired prompt optimization loop for SDD skills. Defines phase signatures, runs eval sets against skill outputs, scores results, suggests prompt modifications, and tracks optimization history. Trigger: When optimizing SDD skill prompts, running eval loops, tuning system messages, or measuring skill output quality.

Standardized evaluation datasets and benchmark suite for SDD artifacts. Defines known-good reference outputs for proposals, specs, designs, and tasks to enable regression testing and quality scoring. Trigger: When building eval datasets for SDD, running quality benchmarks, regression testing skill changes, or comparing model outputs.

Generate CLAUDE.md from codebase analysis using repoforge — hotspots, ownership, hidden coupling, active decisions. No LLM calls needed. Trigger: When user says "generate CLAUDE.md", "analyze codebase for onboarding", or wants automated project context documentation.

Monitors context size and triggers autonomous memory compression mid-session. Detects when the active context is growing large and summarizes older context into engram before it gets compacted and lost. Trigger: When context is large, mid-session, user says "compress", "compact context", or "/memory-compress".

Session memory patterns and toolset profiles for AI coding assistants — persistent context, /remember commands, and mode-based tool access. Trigger: When managing AI context across sessions, implementing /remember patterns, or configuring toolset profiles for different work modes.

CLI-first agent orchestration patterns for headless/non-interactive SDD execution. Provides patterns for running parallel apply, progress reporting to stdout, and batch orchestration without a GUI or interactive user. Trigger: When running agents headlessly, in CI, or user says "headless", "cli orchestration", "batch agents", "/agent-orchestrator".

Renders SDD task lists as a text-based Kanban board with TODO, IN PROGRESS, and DONE columns. Reads task state from tasks.md (openspec) or engram and presents a visual board in the terminal. Trigger: When user says "kanban", "board", "task board", "show board", or "/sdd-kanban".

Automated git worktree workflows — isolate each task in its own worktree, work in parallel, and auto-create PRs. Trigger: When running parallel tasks with git worktrees, automating PR creation from worktrees, or configuring wave-executor for isolated execution.

Explore and investigate ideas before committing to a change. Trigger: When the orchestrator launches you to think through a feature, investigate the codebase, or clarify requirements.

Detects project dependencies and suggests or installs relevant official skills — maps package.json, pyproject.toml, and other manifests to matching skills. Trigger: When onboarding to a project, user says "install skills", "suggest skills", "what skills do I need", or when setting up a new project workspace.

Converts SKILL.md files to other AI assistant formats — .cursorrules, .windsurfrules, AGENTS.md, and raw markdown. Trigger: When user wants to export skills for Cursor, Windsurf, or other AI tools, says "export skill", "convert to cursorrules", or "generate AGENTS.md".

Generates a SKILL.md from documentation URLs, markdown files, or raw text — extracts patterns, rules, examples, and anti-patterns automatically. Trigger: When user wants to create a skill from docs, says "generate skill from", "skill from docs", "convert docs to skill", or provides a URL/file to turn into a skill.

Break down a change into an implementation task checklist. Trigger: When the orchestrator launches you to create or update the task breakdown for a change.

Async notification patterns for SDD workflows via Claude Code Channels (webhooks, Telegram, Discord). Push events when sub-agents finish, verify finds issues, or builds fail. Trigger: When configuring notifications for SDD, user says "notify me", "channels", "webhook", "alert when done", or invokes /channels-setup.

Auto-persisted hot context cache for fast session recovery. Captures active task, recent decisions, and working files at session end, loads them first at session start. Trigger: Session start (auto-load), session end (auto-persist), user says "hot context", "session cache", "what was I working on", or invokes /hot.

Token savings analytics combining RTK gain data with cost-tracking to show real savings per session alongside dollar costs. Gracefully skips when RTK is not installed. Trigger: When user asks about token savings, says "rtk gain", "how much did I save", "token efficiency", or invokes /token-savings.

Six concrete anti-patterns that AI agents consistently produce. Enforces discipline rules to prevent silent fallbacks, catch-all error handling, evergreen tests, hardcoded lookups, skipped TDD, and premature debug log removal. Trigger: When writing or reviewing AI-generated code, during code review, or when applying sdd-apply.

Designing Data-Intensive Applications (DDIA) decision frameworks for architecture decisions. Quick-reference matrices for storage engines, replication, partitioning, and consistency models. Trigger: During sdd-design phase, when making data architecture decisions, or when choosing between storage strategies, replication topologies, or consistency guarantees.

Implement tasks from the change, writing actual code following the specs and design. Trigger: When the orchestrator launches you to implement one or more tasks from a change.

Auto-generates CLAUDE.md capturing project knowledge, flags irreversible decisions (one-way-door), and generates LESSONS.md retrospectives. Trigger: When onboarding to a project, after major changes, user says "generate CLAUDE.md", "capture learnings", "one-way door", or "retrospective".

Runtime safety protocol for detecting and killing runaway SDD sub-agents. Monitors token consumption, tool calls, duration, and consecutive failures. Trigger: When configuring sub-agent safety, discussing runaway agents, or user mentions circuit breaker, runaway, agent limits.

Post-task learning capture workflow that documents what was learned, appends to learnings.md, and suggests CLAUDE.md improvements. Trigger: After completing a significant task, when user says "compound", "what did we learn", "document learnings", or invokes /compound.

Semantic compaction of completed SDD apply sessions. Summarizes closed tasks into a single "Completed Work" block, discarding step-by-step implementation noise while preserving decisions, file paths, and risks. Run after sdd-archive to reclaim context. Trigger: /sdd-compact, or suggested automatically after sdd-archive.

Analyzes a session summary and extracts reusable patterns as proposed SKILL.md drafts. Trigger: /analyze-session, or user says "extract patterns", "make this a skill", "what can we reuse from this session".

Comprehensive LLM evaluation framework with automated metrics, LLM-as-Judge, and RAG evaluation. Trigger: When evaluating LLM outputs, building eval pipelines, comparing models, or measuring RAG quality.

Advanced prompt engineering patterns for production LLM applications. Trigger: When designing prompts, system messages, structured output, or optimizing token usage.

Autonomous try-measure-keep/revert loop for hypothesis-driven code improvements. Runs N experiments, measures each via tests/benchmarks, keeps improvements, reverts regressions. Trigger: When --experiment flag is used with /sdd-apply, or user says "experiment", "try variations", "hypothesis loop".

Security scanner for AI agent skills — detects credential theft, code injection, data exfiltration, and scope escape before installation. Trigger: When installing skills, reviewing skill files, user mentions skill security, or before running `javi-forge plugin add`.