Claude Code Routines

Original Article ↗ Hacker News Discussion ↗

Anthropic’s new “Claude Code Routines” feature—scheduled and event‑driven automations for its Claude Code coding assistant—lands in a climate of growing frustration over rate limits, model “nerfs,” and opaque terms of use. Commenters see potential in letting Claude automatically triage GitHub issues, summarize activity, or maintain code, but worry that these tools deepen vendor lock‑in, are brittle or flaky in practice, and may be withdrawn or degraded over time. Many argue it’s safer to keep core workflows in self‑hosted or open tooling and treat proprietary agent features as interchangeable conveniences rather than foundations.

Feature & Use Cases

  • Routines let Claude Code run tasks on a schedule, via callbacks, or on GitHub events.
  • Users report successful workflows: PR review, Slack/email/GitHub digests, feedback triage, simple automation around repos.
  • Some see it as Anthropic absorbing “OpenClaw-style” cron + hooks; others liken it to n8n / GitHub Actions but LLM-driven.
  • Several say it’s easy to replicate with cron + scripts, so the feature is more about convenience and hosting than raw capability.

Usage Limits, Pricing & ToS Ambiguity

  • Strong confusion around what’s allowed on the fixed-price subscription:
    • Is claude -p allowed in scripts, bots, IDEs, or only direct human use?
    • When does a personal script become a “third‑party harness”?
  • Reports of accounts banned for scripted CLI use, with little recourse.
  • Routines on Max include a small number of “free” runs per day, then bill per-token, which some see as constraining and opaque.
  • Many perceive shifting limits mid‑subscription as bait‑and‑switch; others see it as a compute‑capacity reaction.

Model Quality, Context & “Nerfing”

  • Multiple users feel Claude (especially Opus) has become less reliable, more verbose, more error‑prone in coding.
  • Others still find it excellent, suggesting possible A/B tests, routing differences, or expectation drift.
  • The 1M-token context is widely blamed for token bloat, higher costs, and quality regressions; people manually cap context to ~200k.
  • There’s debate whether 1M vs 200k context variants differ in quality below 200k tokens; outcome is unclear.

Lock‑In, Platform Strategy & Trust

  • Many view Routines as another step toward vendor lock‑in and “AI cloud” platform economics, not just model access.
  • Strong reluctance to depend on opaque, changing features (Routines, Skills, Cowork) that could be nerfed, sunset, or repriced.
  • Comparisons to cloud lock‑in (AWS Lambda et al.); several prefer keeping orchestration under their own control.

Alternatives, Reliability & Broader Sentiment

  • Frequent mentions of alternatives (OpenClaw, GitHub Agentic Workflows, Codex, local/open models, custom orchestrators).
  • Complaints that Anthropic ships overlapping, sometimes buggy features while core issues (context bloat, CLI regressions, flaky scheduling) persist.
  • Growing frustration with rapid “feature velocity,” unclear policies, and perceived enshittification, even from long‑time fans.