A Hacker News thread posted on June 17th asked a deceptively simple question: "What are your best Claude hacks?" The post, which garnered modest engagement with a score of just 4 points, turned into a goldmine of practical workflow tips from developers who've spent serious time pushing Anthropic's AI assistant beyond its defaults. Rather than surface-level prompting tricks, the top responses dug into architectural patterns for working with Claude at scale—approaches that could save hours of frustration for anyone building automated pipelines or complex multi-step projects.

Sub-Agents Keep Context Clean

The most upvoted technique involves treating Claude like a software engineering project itself. One contributor described using sub-agents to handle discrete tasks—code reviews, research queries, plan validation—while keeping the parent conversation's context uncluttered. Instead of asking one massive Claude instance to juggle everything, power users spawn focused agents for specific jobs, then synthesize results back into the main thread. This mirrors how senior engineers modularize monolithic applications: small parts that work in isolation, verified independently, then integrated into larger systems.

Verification Loops With /goal

Several commenters emphasized the importance of building verification checkpoints into prompts. When tasks have measurable outcomes—a passing test suite, a screenshot matching expected pixels, a specific output string—using /goal with explicit acceptance criteria dramatically improves reliability. The strategy works because it forces Claude to self-check before declaring victory. One user noted that when "there's ever something you can have verification on," leaning into automated checks eliminates the back-and-forth of correcting hallucinations or drifted outputs.

Skills as Persistent Memory

The final major hack involves teaching Claude to remember solutions to recurring problems. When Claude repeatedly stumbles on a specific task—say, running Playwright tests or authenticating with a NAS—the solution isn't endless retuning; it's writing a skill. These persistent instructions encode the paths that worked before, allowing Claude to follow proven routes rather than reinventing wheels and burning tokens on failed attempts. Users reported this was especially valuable for infrastructure-adjacent tasks where context windows get tight and consistency matters more than creativity.

Key Takeaways

  • Modularize your workflow: spawn focused sub-agents instead of dumping everything into one conversation
  • Build verification checkpoints wherever possible—test suites, screenshot comparisons, exact output matching
  • Use /goal with explicit acceptance criteria to reduce hallucination and iteration cycles
  • Write skills for recurring failure points rather than re-explaining context each session