Agentic Coding
Top tech stories today across software, hardware, AI, and product launches. Senior engineer audience — skip rumour churn and pre-announcement leaks. Lead with shipping-now stories and what changed for builders.
Spec-first stops 80% of agent failures, Matt Pocock ships composable skills, harness patterns get a reference stack
1 Min. Lesezeit
Spec-driven development
The number one reason your agent goes sideways is a missing spec.
A new best-practices breakdown argues that writing a detailed markdown specification before you even prompt—covering what gets built, what doesn't, edge cases, and acceptance criteria—prevents roughly 80% of common agentic coding failures [Source: Blink]. The trick is embedding explicit review points inside the spec where the agent must pause and show you the work—after database schema setup, before touching auth—rather than hoping vague instructions stick. Combine this with plan mode whenever you're changing three or more files, and you catch scope creep before it's implemented.
Your spec is the guardrail your agent actually respects.
Matt Pocock's skills
Matt Pocock just open-sourced his entire .claude directory.
The new skills repo gives you composable tools like /grill-me for pre-coding alignment, /to-prd for converting conversations into PRD documents, and /tdd for test-driven feedback loops [Source: GitHub]. Install with a single npx command, pick the skills you want, and run /setup-matt-pocock-skills to configure your issue tracker and docs structure. These aren't vibe-coding tricks—they're production patterns from decades of engineering experience, built to slot into your existing Claude Code or Cursor workflow.
Grab /improve-codebase-architecture before your next refactor.
Context and escape valves
You're probably letting your agent touch code it shouldn't.
The same Blink guide recommends defining explicit escape valves: tell the agent which code areas require human approval before modification and what to do when uncertain [Source: Blink]. Pair this with aggressive context hygiene—use /clear at task boundaries, /compact during long sessions, and keep your CLAUDE.md concise with stack decisions and out-of-scope zones. Writing tests first as acceptance criteria also gives your agent a built-in verification loop.
Guardrails beat cleanup every time.
Agentic Coding Best Practices: Spec-Driven Development, Human ...17 hours ago ... Guideagentic codingclaude codecursor. Agentic Coding Best Practices: Spec ... Commands the agent would otherwise forget (test scripts, environment setup).blink.new

Spec-driven development is the foundation for successful agentic coding workflows. Write a detailed specification in markdown format before prompting the agent, including what gets built, what doesn't, edge cases, and acceptance criteria — this practice prevents about 80% of common agentic coding failures. Set explicit review points within the spec where the agent must stop and show you the work before continuing, such as after database schema setup or before touching authentication layers, rather than relying on vague instructions. Use plan mode before execution when three or more files will be changed, treating the agent's proposed plan like a code review to catch scope creep and wrong assumptions before they're implemented. Manage context aggressively by using /clear at task boundaries between features and /compact during long sessions, while maintaining a concise CLAUDE.md file with stack decisions, out-of-scope areas, and commands the agent needs to remember. Write tests first as acceptance criteria, start with small vertical slices of features rather than building the entire system at once, and define explicit escape valves by telling the agent which code areas require human approval before modification and what to do when uncertain.
skills/README.md at main · mattpocock/skills - GitHub20 hours ago ... I built these skills as a way to fix common failure modes I see with Claude Code, Codex, and other coding agents. ... best debugging practices into a simple loop.github.com
The content provides practical agent skills and workflow best practices directly aligned with your interests. Matt Pocock's engineering-focused skills address core challenges in agentic coding: alignment between you and the agent through /grill-me and /grill-with-docs (which help before coding starts), reducing agent verbosity via shared domain language in CONTEXT.md documents, and improving code quality through feedback loops like /tdd for test-driven development and /diagnose for debugging. Key skills include /to-prd for converting conversations into PRD documents, /improve-codebase-architecture for maintaining system design quality, and /zoom-out for architectural context. The setup is straightforward—run npx skills@latest add mattpocock/skills, select your desired skills and coding agent (works with Claude Code and others), then run /setup-matt-pocock-skills to configure your issue tracker and documentation structure. These skills are designed as composable, adaptable tools based on decades of engineering experience rather than "vibe coding," making them suitable for solo founders building SaaS applications who want production-quality results with AI agents.