workflow

As a conversation grows longer, the model gradually ignores earlier instructions and grows unreliable, even without hitting the <a href="https://claude-me.com/en/glossary/core-concepts/token/" target="_blank" rel="noopener">Token</a> limit — a gradual decline, not a sudden cliff.

The maximum amount of text Claude can "see" at once within a single conversation, measured in tokens — covering the system prompt, conversation history, uploaded files, and the response being generated. It's not how much Claude permanently remembers; it's how much working-memory space this conversation currently has.

When writing instructions or a <a href="/en/glossary/workflow/skill/">Skill</a>, matching how specific the guidance should be to how error-tolerant the task actually is — deciding whether Claude gets an exact sequence to follow or just a general direction to judge for itself.

Claude Code's built-in environment health-check command, which diagnoses issues with installation, configuration, permission rules, and MCP connections, and can help fix them once you confirm.

The coordination layer that decides who does what and when, and how results get passed along — a separate mechanism from content generation itself, responsible for dispatching and integrating work.

Splitting documentation into a main file plus several extension files, where the main file holds only a high-level overview and links pointing to detail, and the extended content only gets read when it's actually needed — so a large body of reference material doesn't consume context space upfront.

A reusable bundle of instructions and scripts that lets Claude follow an established workflow for a specific task type, instead of having the process re-explained from scratch every time.

A separate Claude instance the main session spawns to handle one specific job, with its own <a href="https://claudecowork-me.com/en/glossary/core-concepts/context-window-workplace/" target="_blank" rel="noopener">Context Window</a> and tool permissions, reporting back only the final result rather than flooding the main conversation with every intermediate step.

A YAML-formatted <a href="https://crypto-bible.com/en/glossary/blockchain-fundamentals/block/" target="_blank" rel="noopener">Block</a> wrapped in --- markers at the very top of a Markdown file, used to describe the file's metadata — the part of a <a href="/en/glossary/workflow/skill/">Skill</a> that always loads first in the three-layer <a href="/en/glossary/workflow/progressive-disclosure/">Progressive Disclosure</a> model.