Claude Code Routines — Helicopter View

The Problem

One Session, One Task, Manual Prompting

You prompt each step by hand. Context dies on every break (5-min cache TTL). Multi-step workflows mean either babysitting a single session or trusting a bash script with no reasoning. The agent forgets what comes next.

What Are Routines

Saved Agentic Workflows With Branching Reasoning

A Routine is a defined sequence of agentic steps Claude Code executes without per-step prompting. Each step is a prompt-plus-tool-call pattern with conditional branching.

Unlike a bash script, the agent still reasons at each step. Unlike a one-shot prompt, the structure is reusable and pauseable.

Three components: trigger (slash command, file change, manual), steps (prompt + allowed tools + on_fail branch), completion condition.

Mental Model

Agent as Workflow Executor

Old Claude Code: conversation partner you nudge step-by-step. New Claude Code: workflow executor you hand a shape to. The reasoning stays. The pushing-forward disappears.

Routine Shape

YAML Definition

name: blog-post-check trigger: manual steps: - prompt: "Lint MDX" tools: [read, edit] - prompt: "Run build" tools: [bash] on_fail: "Fix & retry" - prompt: "Stage & commit" requires_approval: true

~90s vs 6-8 min manual.

Multi-Session

Parallel Contexts in One App

Two-to-three sessions at once. One focused. One background Routine. One quick lookup that does not pollute the main thread. Shared filesystem means serialise file edits, not attention.

Session A|| Session B|| Session C

1-Hour Prompt Cache

Cache Survives Your Coffee Break

v2.1.108 adds ENABLE_PROMPT_CACHING_1H=true. Default 5-min TTL went cold on every meeting; 1h TTL survives standups and reviews. Hit rate jumped from ~60% to ~85% in my use.

~/.zshrc export ENABLE_PROMPT_CACHING_1H=true

LSP-over-Grep Hooks

Query the Language Server, Not the Filesystem

Instead of grepping, Claude Code queries your project's LSP for definitions, references, and types. Claimed 80% token savings (your mileage varies by LSP quality). Answers about code structure are more accurate, fewer "let me search for that" steps.

Before vs After Routines

What Actually Changed

Dimension	Old Claude Code	With Routines + Multi-Session
Workflow	Manual prompt per step	Agent executes defined shape
Sessions	One at a time	2-3 parallel contexts
Cache TTL	5 minutes	60 minutes (opt-in flag)
Code queries	Grep + file reads	LSP (definitions, types, refs)
Failure handling	Stops, waits for you	on_fail branch, retries once
Destructive ops	Always asks	requires_approval gate per step

Key Decisions

How to Adopt Without Over-Investing

Start with the built-in Routines: review-and-commit, test-then-fix, refactor-with-verify
Encode workflows you run daily, not one-off tasks
Put requires_approval on any destructive step (commits, pushes, deletions)
Use multi-session for long-running tasks, not to multitask yourself
Enable 1h caching globally in your shell profile
Skip v2.1.107 — go straight from 106 to 108

Invoke

Slash command File change Manual trigger

Status

Research Preview Desktop v2.1.108

Still Rough

Known Limitations

Context degrades after ~2h active chat
Routine failures give brief errors, not full reasoning traces
No Routine sharing, versioning, or plugins yet
Integrated terminal is not a full replacement for tmux
v2.1.107 had tool-call regressions (fixed in 108)
Longer cache TTL means stale CLAUDE.md sticks around longer

What Else Shipped

The Desktop Redesign at a Glance

Integrated terminal inside the app (common flows, not all tmux tricks)
Inline file editing without leaving the session
HTML and PDF preview panels
Multi-session switcher with shared filesystem guardrails
LSP-over-Grep hooks for symbol-aware code queries
1-hour prompt caching flag behind an env var

The signal

Claude Code is shifting from agent-as-conversation-partner to agent-as-workflow-executor. Routines are the infrastructure; multi-session and caching are what make them practical for a full working day.

Components

Routines Multi-Session 1h Cache TTL LSP Hooks Inline Editor HTML/PDF Preview