3b-forge

Socratic interview to crystallize vague requirements into clear specifications. Cross-agent: Claude Code, Codex, Gemini CLI.

Discover, sync, and publish blog posts from 3B knowledge base. Runs dry-run sync to find publishable posts, lets user select which to publish, syncs English versions, expands them into narrative blog posts, then translates to Korean using /translate-ko workflow. Use when user says "publish blog", "sync posts", "new blog posts", or "blog publish".

Audit and repair all symlinks from 3B to global config profiles (~/.claude, ~/.claude-work) and project repositories. Use this skill whenever the user says "check symlinks", "symlink audit", "verify symlinks", "fix symlinks", "symlink health", or "check links". Also trigger when the user reports settings not propagating, config drift between profiles, or after Claude Code UI updates or plugin reinstalls — these commonly overwrite symlinks with regular files via atomic rename.

Automate edX course study sessions using Chrome MCP browser tools. Claude reads page content via screenshots, analyzes quiz questions and presents answers in batch tables, navigates between pages, and creates study notes progressively. User handles all iframe interactions (selecting answers, clicking Submit, scrolling quiz pages). Two modes: study (notes + explanations) and exam (fast answers only). Use when user says "edx study", "let's study", "continue studying", "start exam", "exam mode", "practice exam", "course exam", or "edx [course-name]".

Intake, analyze, and process research papers (PDF or arXiv URL) into the 3B knowledge system. Extracts academic metadata, generates structured analysis, identifies 3B/Claude applicability, creates knowledge entries, and drafts blog + LinkedIn posts. Maintains a searchable paper library index. Use when user says "research paper", "analyze paper", "read this paper", "paper review", shares an arXiv link/PDF path, or wants to process a research paper.

Aggregate journal entries into higher-level summaries using clean-slate rollup strategy. Combines daily entries into weekly, weekly into monthly, monthly into quarterly, quarterly into yearly. Cleans ephemeral levels after aggregation. Use when user says "rollup", "aggregate journals", "weekly summary", "monthly rollup", or "summarize this week/month".

Translate English content to natural Korean. Supports blog posts, PR descriptions, documentation, READMEs, and general text. Uses voice calibration from writing examples and a 4-step workflow (Analyze, Translate, Rewrite, QA). Use when user says "translate to Korean", "Korean translation", "translate this", "ko version", or "한국어로".

Create GitHub issues with conventional commit title format, bilingual body (English + Korean), proper labels, assignee, and project integration. Enforces Issue Type selection, required labels, and project assignment. Use when user says "create issue", "new issue", "open issue", "file issue", "report bug", "request feature", or "track task".

Create GitHub pull requests with bilingual notes (English + Korean) and full project integration. Confirms target branch, analyzes commits since divergence, generates structured PR body, assigns user, adds conventional labels, and sets GitHub Project status to "In Review". Use when user says "pr to dev", "pr to prod", "create pr", "open pr", "submit for review", or "push and pr".

Proactive PR code review using 3 specialized agents in parallel. Analyzes PR diff across 7 categories (Security, Code Quality, Performance, Architecture, Test Quality, Maintainability, Deployment Safety) and produces structured findings with severity-based triage. Use when user says "review pr", "proactive review", "code review pr", or "/review-pr". Companion to validate-pr-reviews (reactive). Do NOT use for responding to AI reviewer comments (use validate-pr-reviews instead).

Validate AI-generated PR review comments from Claude, Copilot, and Codex. Fetches unresolved comments, classifies each as valid/invalid/good-to-have/controversial, applies user-approved fixes, and logs agent confusion patterns. Use when user says "validate reviews", "check pr comments", "handle review feedback", "address pr reviews", or "process copilot/claude/codex comments". Do NOT use for writing self-review comments (use add-pr-self-reviews instead).

Add self-review inline comments to GitHub PR diff pages explaining key implementation decisions. Posts comments in Korean (default) or English on specific diff lines. Use when user says "add self review", "self review comments", "annotate pr", "explain my changes", or "add review notes". Typically run after /validate-pr-reviews iterations are complete.

Resolve requirement, intent, or strategic uncertainty through structured hypothesis-driven questioning. Three auto-detected modes: vague (ambiguous requirements to concrete specs), unknown (strategic blind spots to 4-quadrant playbook), metamedium (content vs form reframing to leverage analysis). ALWAYS uses AskUserQuestion. Trigger on "clarify", "scope this", "spec this out", "what am I missing", "blind spots", "content vs form", "diminishing returns", "make this clearer", "assumption check", "요구사항 명확히", "뭘 놓치고 있지", "뭘 모르는지 모르겠어", "같은 방식이 안 먹혀", "관점 전환", "전략 점검", "내용 vs 형식". Standalone or sub-workflow for other skills.

Audit documentation for stale references, broken links, missing cross-references, and outdated content. Scans markdown files, validates frontmatter, checks relative links, and reports parity gaps. Use when user says "audit docs", "doc audit", "check stale links", "find broken references", "documentation health", or "validate docs".

any input (code, docs, papers, images) - knowledge graph - clustered communities - HTML + JSON + audit report

Investigate a bug, issue, or question by running parallel searches across codebase, git history, docs, and external sources. Produces a consolidated analysis with root cause hypothesis and recommended fix. Use when user says "investigate", "dig into", "root cause", "why is this happening", "trace this bug", or "figure out why".

Track repeating task patterns after completing non-trivial work. Normalizes tasks to kebab-case labels in `~/.claude/task-tracker.json`, bumps counts, and suggests skill/hook/command automation when thresholds are met. Trigger after finishing any multi-step task that could be templated, or when the user says "track this pattern", "tracker check", or "suggest automation". Also invoked automatically by `/wrap` Step 4.7.