Shift Log: Slash Commands That Cut Claude Code Delay

Keep the shift state visible before changing anything

Experienced Claude Code users lose time less often from typing than from losing track of the session. Start a meaningful task with /status to inspect the active model, workspace state, and connected services. Use /context when a long investigation starts producing vague answers or repeated questions. It shows how much context is in play and helps you decide whether to continue, compact, or reset.

/cost is useful at the end of an expensive investigation, but it is also valuable during one. Check it after a large repository scan, a test-failure loop, or a multi-file refactor. The command gives you an operational signal before you commit to another broad pass.

Compact deliberately when the investigation has changed shape

/compact is more effective when you treat it as a checkpoint rather than an emergency button. Before compacting, ask Claude Code to summarize the current diagnosis, files changed, commands run, failed hypotheses, and remaining acceptance criteria. Then run /compact. The retained context is more useful when the important state has been written down explicitly.

This is especially helpful after switching from discovery to implementation. A repository tour, dependency trace, and incident timeline are valuable during diagnosis but can become noise once the fix is clear. Compacting at that boundary keeps the next phase focused without discarding the working conclusion.

Use clear for hard boundaries between unrelated tasks

/clear is the right command when the next task should not inherit assumptions from the previous one. Use it between unrelated tickets, after abandoning a misleading hypothesis, or before reproducing a bug from a clean prompt. Clearing is also a practical way to prevent old file names, requirements, or temporary decisions from influencing a new change.

For recurring work, pair /clear with a concise task brief that states the target behavior, constraints, verification command, and files that are in scope. This produces a more repeatable handoff than asking the model to ignore everything it already knows.

Switch models at the point of diminishing returns

/model makes model selection part of the workflow instead of a permanent session setting. Use a faster model for repository reconnaissance, command interpretation, and small edits, then move to a stronger model when the task requires architectural judgment or a delicate migration. For example, Claude Haiku 4.5 can handle a quick first pass, while Claude Opus 5.5 may be more appropriate for a difficult cross-module design review.

The useful habit is to switch based on the next decision, not on the prestige of the model. A capable model does not remove the need to provide tests, constraints, or a precise stopping condition.

Inspect memory and permissions before automating repetition

/memory shows the instruction files Claude Code can use for persistent project guidance. Review those files when behavior seems strangely consistent across sessions, especially after changing build commands, style rules, or test conventions. Keep durable guidance short and verifiable; a large memory file becomes another source of stale context.

/permissions is worth checking before a batch operation or when a command unexpectedly pauses for approval. Tight permissions make risky actions visible, while carefully chosen approvals reduce interruption during trusted local workflows. Revisit them when moving between repositories with different security expectations.

Make custom commands carry the runbook

Project and personal custom slash commands are where experienced users get compounding returns. Put the repeatable sequence for tasks such as release checks, dependency audits, incident summaries, or focused test runs into a command file. The command should state inputs, required tools, output format, and verification steps. Keep the procedure narrow enough that a failed command identifies one problem instead of creating an opaque automation layer.

When the team uses multiple model providers, Relay365 can fit this workflow as a pay-as-you-go relay desk for Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok, and other model families. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients through https://api.api365relay.com, with OpenAI clients adding /v1. Create a Relay365 account when you need one billing and routing point across these tools. Members can also earn a referral bounty when invited users pay.

Close every shift with a recoverable record

Before ending a session, run /cost, record the current state with a short summary, and leave the repository passing the relevant checks or clearly marked with the remaining failure. The fastest future session is usually the one that starts with an accurate handoff, not the one that begins with the longest prompt.