# Navigating Claude Code: The Context Window Tax > [!question] Lingering Questions > - The article treats context size as a major cause of quality loss without isolating it from other causes (task complexity, model choice) — the deep-read flags this directly as a limitation. > - Token-accounting examples and context-window sizes given are 2026 snapshots — recheck against current limits before using any specific number. > [!tip] Article Ideas > - "Filter before Claude reads" is a concrete, testable practice — could anchor a short piece on designing CLI tool output for agent consumption specifically (verbose-by-default tooling is expensive when an LLM is the reader). > [!link] Cross Reference > - Part 4 of Oleg Efimov's 10-article HackerNoon series; see [[3. Resources/Literature Notes/Navigating Claude Code - The Full Workflow]] for series context. > - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — same "the schema/context accompanies every request, not just the ones that use it" mechanism, from an MCP-specific source > [!quote] Vocabulary > - **Context window tax** — the article's term for the recurring per-turn cost of accumulated conversation, tool output, tool schemas, and instructions, paid whether or not that content is relevant to the current turn ## Bibliographic capsule Article, HackerNoon, 2026-08-04. Fourth installment of Oleg Efimov's Claude Code series. ## Author's thesis (Adler Q1+Q2) **What is the article about?** Why accumulated conversation, tool output, and instructions impose a recurring cost on every turn of a Claude Code session, and how to budget against it. **What is being argued?** Healthy long sessions require active context budgeting: clear unrelated work rather than carrying it forward, compact with a stated focus when continuity matters, and filter verbose output before the model reads it — because every previous turn's content is re-read, not just referenced, on each subsequent turn. ## Key arguments (≤3) 1. **Clear unrelated work instead of carrying old context forward** — starting a new unrelated task in the same session pays for the old task's context on every subsequent turn. 2. **Compact with a stated focus when continuity matters** — an unfocused compaction risks losing the specific thread that's actually needed. 3. **Filter verbose output and disable unused tools before the model reads them** — tool schemas and command output both count against the same per-turn budget whether or not they're used. ## Claim atoms ### Clear unrelated work instead of carrying old context forward > [!quote] ID:rw_01kz7st5ya2snaendt9476tw47 > context window tax Where/why: names the article's core concept. ### Compact with a stated focus when continuity matters > [!quote] ID:rw_01kz7st6jznhmfj02w2y6mn26g > Use /clear between unrelated tasks Where/why: the article's practical mitigation for task-switching cost. ### Filter verbose command output and disable unused tools before the model reads them > [!quote] ID:rw_01kz7st7d8hdajab4jws314yr4 > Filter before Claude reads Where/why: extends the budgeting principle from conversation history to tool output specifically. ## Permanent candidates - **Standing capability and accumulated history both charge a fixed per-turn cost, not a per-use cost** — the article's tool/output-filtering argument is the same shape as an existing vault claim about MCP tool schemas; connect rather than duplicate (see Links). ## Processing checklist - [x] Checked for an existing Permanent note covering this article's core mechanism — found one from an MCP-specific source, linked rather than duplicated as a new note - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]], [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] ## Links - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — a developer-facing instance of the standing-cost-per-turn pattern already tracked in this MOC - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — session-management practice for agentic coding workflows - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — because evidence-for; a second, independent source (Claude Code session context vs. MCP tool schemas) confirming the same "standing content is re-paid every turn" mechanism - Metadata: [[3. Resources/Readwise/Articles/Navigating Claude Code The Context Window Tax|Navigating Claude Code - The Context Window Tax]]