Maximizing the value of your Claude Code sessions

Original Article ↗ Hacker News Discussion ↗

Anthropic’s guidance on “maximizing” Claude Code sessions prompts mixed reactions from developers trying to control token usage, cache behavior, and costs in an opaque, fast-changing tool. Many welcome concrete tips on prompt caching, context management and subagents, but others argue these optimizations should be handled automatically by the product rather than offloaded to users who now bear direct per-task costs. The exchange highlights broader unease about vendor incentives, frequent behavioral changes in hosted AI services, and comparisons with both cheaper competitors and more transparent, self-hosted setups.

Perceived Inefficiency & Bugs in Claude Code

  • Some users report Claude Code is slower and less targeted than alternatives (e.g., Codex/Copilot), over-reading files and directories instead of staying focused on the given file.
  • Unexpected prompt cache rewrites and high cache-write counts are driving very large bills for some, even when they believe they’re following best practices.
  • Multiple GitHub issues are cited (cache bugs, /clear bleeding into next session, new sessions not hitting full cache), leading to frustration and distrust in the tooling.
  • Desktop app @-file search is reported as buggy or inferior to CLI.

Token Costs, Cache Behavior & Incentives

  • Many comments focus on prompt caching: TTL differences (5 minutes vs 1 hour), cache-busting from changing model/effort, using /compact, and confusion about when caches reset.
  • Some argue Anthropic is aligned with users in wanting fewer tokens (compute is expensive; subscriptions cap usage); others counter that enterprise per-token billing incentivizes more token usage.
  • There’s skepticism around “cache-busting” behavior and whether optimizations favor user savings or provider profit.

Product Design vs. “You’re Holding It Wrong”

  • Strong sentiment that the blog post offloads complexity (context management, compaction, cache TTL, effort levels) onto users instead of building smarter defaults.
  • Several see the guide as the AI equivalent of “you’re holding it wrong”: if misuse is common, they argue it’s a product design failure.
  • Others defend the post as normal “how to use a powerful tool efficiently,” analogous to cost-optimization guides for AWS or databases.

Comparisons to Other LLMs & Local Harnesses

  • Some report better cost/performance using other LLMs (OpenAI, DeepSeek, Kimi, Qwen) or custom harnesses with version-pinned configs and local / cloud-hosted models.
  • Usage limits and perceived higher cost/latency of Claude Code are frequent reasons cited for preferring competitors.

Workflows, Commands & Skills

  • Users share advanced practices: short sessions, frequent /clear or /compact, /handoff to new sessions or other models, “docset-driven” development, and external planning documents.
  • Debate over whether @-mentioning large files is good (cacheable, always present) or bad (forces full reads).
  • Confusion and annoyance that changing “effort level” busts cache; speculation it’s implemented via hidden system prompts.

Broader Concerns About Opaque, Changing Tools

  • Some want inspectable, stable tools and resent rapid, opaque changes in hosted AI services.
  • Others accept churn as inevitable for fast-evolving agentic tooling, arguing expectations of stability are unrealistic right now.