hivellm

Create structured analyses with numbered findings, execution plans, and task materialization

Author a rulebook task spec interactively — research, draft, ask the user clarifying questions, confirm, then create the tasks in rulebook ready for /rulebook-driver. Use when the user wants to plan/spec a feature before implementing.

Spec-driven task management for features and breaking changes with OpenSpec-compatible format

Tool: Anthropic Claude Code CLI (`npm install -g @anthropic-ai/claude-code`)

Behavioral guidelines to reduce common LLM coding mistakes — overcomplication, sloppy refactors, hidden assumptions, weak goals. Use when writing, reviewing, or refactoring code. Auto-applies; invoke explicitly via /karpathy-guidelines or 'follow karpathy discipline'.

Accessibility audit for WCAG compliance and usability

Design or review API endpoints (REST/GraphQL)

Architecture review, design decisions, and ADR writing

Detect and fix documentation violations across C++ compiler and TML library code. Scans for missing ///, undocumented public items, wrong comment styles, and auto-fixes using parallel agents. Use when the user says "fix docs", "audit docs", "documenta", "corrige docs", or wants to improve documentation coverage.

Deep knowledge of TML compiler internals. Inject before working on compiler C++ code. Maps the full pipeline with key functions, types, and cross-subsystem boundaries.

Deep knowledge of TML standard library structure, module dependencies, key types, and C runtime mapping. Inject before working on library code.

Detect which tests are affected by recent code changes. Use when the user wants to know what tests to run after making changes.

List build artifacts with sizes and ages. Use when the user asks about build outputs, compiled binaries, or disk usage.

Build the TML compiler from C++ sources. Use when the user says "build", "compila", "rebuild", or needs to recompile the compiler after C++ changes.

Detect what changed and build only the necessary target. Use when the user says "build smart", "build inteligente", or wants an optimized build.

Invalidate compilation cache for specific source files. Use when cached results seem stale or after external changes to source files.

Create structured analyses with numbered findings, execution plans, and task materialization

Terse Conventional Commits generator. Subject ≤50 chars (hard cap 72), body only when 'why' isn't obvious, no filler. Use when user says 'write a commit', 'generate commit', 'commit message', or invokes /rulebook-terse-commit. Independent of the base rulebook-terse mode.

Ultra-compressed code review comments. One line per finding: location, problem, fix. Use when user says 'review this PR', 'code review', 'review the diff', or invokes /rulebook-terse-review. Independent of the base rulebook-terse mode.

Output-verbosity compression. Cuts response tokens ~40-70% without losing technical accuracy. Four intensity levels (off/brief/terse/ultra) aligned with Rulebook's agent-tier system. Use when user says 'terse mode', 'be terse', 'less tokens please', or invokes /rulebook-terse. Auto-activates via SessionStart hook with tier-aware default.

Author a rulebook task spec interactively — research, draft, ask the user clarifying questions, confirm, then create the tasks in rulebook ready for /rulebook-driver. Use when the user wants to plan/spec a feature before implementing.

Save session state to .rulebook/handoff/_pending.md for cross-session continuity

Fetch the audit envelope for a past Cortex turn or query. Returns caller, intent, scope, per-field counts, latency, cache hit/miss, and a structured trace of what each retrieval lane returned. Use this when the user asks "what did Cortex do?", "why did the model see X?", or wants to debug retrieval quality on a specific turn.

Refresh the system-prompt context bundle for the current task. Pulls active laws, recent decisions, similar past turns, and relevant snippets via the Cortex pre-thinking pipeline. Use this when the user asks to "load context", "remind me what we know about X", "before I change Y", or whenever the conversation needs grounding against Cortex's institutional memory.

Surface active laws in the current scope (repo + topics). Returns law id, severity, title, and detector for each rule that applies. Use this when the user asks "what laws apply here?", "is X allowed?", "any rules I should know?", or before suggesting an action that might trip a critical law.

Behavioral guidelines to reduce common LLM coding mistakes — overcomplication, sloppy refactors, hidden assumptions, weak goals. Use when writing, reviewing, or refactoring code. Auto-applies; invoke explicitly via /karpathy-guidelines or 'follow karpathy discipline'.

Create structured analyses with numbered findings, execution plans, and task materialization