Agentic Coding

Matt Pocock's agent skills collection addresses key failure modes in AI-assisted coding through structured workflows. The core approach centers on four problems: misalignment between developer intent and agent output (fixed via /grill-me and /grill-with-docs skills that establish detailed requirements upfront), agent verbosity (solved by building a shared domain language documented in CONTEXT.md and ADRs to reduce token usage and improve codebase navigation), code quality issues (addressed through feedback loops including test-driven development with /tdd and systematic debugging via /diagnosing-bugs), and architectural degradation (tackled with /improve-codebase-architecture to prevent code entropy). The skills are composable, model-agnostic, and emphasize small deliberate steps with rapid feedback cycles. Key techniques include grilling sessions to align on domain terminology, using shared language glossaries to enable concise agent-developer communication, and red-green-refactor loops for consistent code quality. These practices prioritize software engineering fundamentals over rapid iteration, targeting developers using Claude Code and similar agents daily.

Agent Skills are structured instruction folders that help AI agents like Claude Code, Cursor, and Cline work more efficiently with specific frameworks and tools. Supabase's Agent Skills repository provides ready-to-use skills covering Supabase development and Postgres best practices that can be installed as Claude Code plugins or integrated with other compatible AI agents. The Supabase skill offers comprehensive guidance on database, auth, edge functions, and client library integration, while the postgres-best-practices skill provides performance optimization guidelines across eight categories prioritized by impact, including query performance, connection management, and schema design—all automatically available to agents once installed to help with common development tasks.

This content provides comprehensive guidance on Claude Code permission management and workflow optimization. It covers five permission modes (Normal, Auto-Accept, Plan, Don't Ask, and Bypass) designed for different development scenarios. For efficient workflow, users can press Shift+Tab to cycle through modes instantly or set a persistent default mode in settings.json. The Auto-Accept mode eliminates permission prompts during file edits, ideal for large refactoring sessions and following well-defined implementation plans. The guide also explains how to use the /permissions command to pre-approve specific tools and patterns, set granular permission rules (allow, ask, deny), and combine permissions with sandboxing for defense-in-depth security. For active development workflows, the recommendation is to use Auto-Accept mode with pre-approved common commands like npm and git while maintaining prompts for system operations, avoiding constant manual approval clicks that defeat productivity purposes.

Claude Code fast mode delivers 2.5x faster Opus 4.6 responses through infrastructure-level priority without sacrificing model quality or reasoning depth. Activated with the /fast command, fast mode maintains identical output quality compared to standard Opus 4.6 but costs more per token. The feature excels during interactive pair programming sessions, rapid iteration cycles, and live debugging where latency directly impacts productivity and focus, though it's less valuable for autonomous background tasks. Fast mode can be combined independently with effort levels and model selection to optimize for specific task requirements—high effort with fast mode for complex architectural work, or fast mode with reduced effort for straightforward refactoring and boilerplate generation.