# Navigating Claude Code: The Full Workflow > [!question] Lingering Questions > - Every specific command name, precedence rule, and cost figure in this series is time-sensitive to a 2026 Claude Code release — the deep-read flags this as the primary limitation across all ten articles and recommends verifying against current docs before acting on any of it. > - The article generalizes from one practitioner's observed workflow with no controlled comparison — how much of the advice survives contact with a different repo size, team, or task mix? > [!tip] Article Ideas > - None distinct to this article — see the series-level note below; the "Full Workflow" piece is the series overview and doesn't carry a standalone angle beyond its sibling articles. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/A Control Is Enforceable Only When the Actor Cannot Reclassify It]] — this article's core distinction ("CLAUDE.md is advisory, hooks are deterministic") is the same enforceable-layer principle already captured from a different source > - This is article 1 of a 10-article HackerNoon series by Oleg Efimov; see the sibling Literature Notes for CLAUDE.md Done Right, The Bare Minimum Setup, The Context Window Tax, Models/Tiers/Effort, Hooks That Guard Your Code, MCP Servers Worth Adding, and Skills That Actually Fire — all six were processed in the same synthesis batch (2026-08-16) and share this article's boilerplate structure and evidence caveats. > [!quote] Vocabulary > - None specific to this article beyond standard Claude Code terminology (CLAUDE.md, hooks, MCP, skills, subagents) — covered per-topic in the sibling notes. ## Bibliographic capsule Article, HackerNoon, 2026-08-04. First installment of Oleg Efimov's ten-article "Navigating Claude Code" series. Practitioner explanation of Claude Code as a layered configuration system, with configuration examples and operational heuristics rather than controlled benchmarks. ## Author's thesis (Adler Q1+Q2) **What is the article about?** How to treat Claude Code as a layered operating system for development work rather than a single tool with one setting to tune. **What is being argued, and how?** Efimov argues that six distinct control layers — persistent project memory (CLAUDE.md), model/effort choice, deterministic hooks, scoped tools, isolated subagent research, and explicit human verification — solve different problems, and conflating them (e.g. treating CLAUDE.md instructions as if they were enforced) causes the failures he's observed in practice. No controlled comparison is offered; the argument rests on configuration examples and stated mechanism ("extra context, tool schemas, or coordination creates cost; narrower routing should reduce it"). ## Key arguments (≤3) 1. **Scope the question before exploration** — unscoped file traversal consumes the main context window; the single most reliable session-extension move is scoping work before delegating exploration. 2. **CLAUDE.md is advisory, hooks are deterministic** — non-inferable project constraints belong in CLAUDE.md; anything that must actually be enforced belongs in a hook, because the model can choose to ignore prose instructions. 3. **Verification is the human's job, not Claude's** — bounding MCP tools, skill activation, and subagent output to the task doesn't remove the need for the developer to check the result. ## Claim atoms ### Scope the question before exploration so file traversal doesn't consume the main context window > [!quote] ID:rw_01kz7ssy8sea97s5kqg13pefx2 > The single most reliable way to extend a session's useful life is to scope work before sending Claude to explore it. Where/why: opening operational recommendation; sets up the series' recurring "narrowest enforceable layer" theme. ### Put non-inferable project constraints in CLAUDE.md, mandatory checks in hooks > [!quote] ID:rw_01kz7ssvs1z0hcbnxpy0na184m > CLAUDE.md is advisory, hooks are deterministic. Where/why: the article's central layering distinction, restated across the whole series with different examples. ### Verification is the developer's responsibility, not Claude's > [!quote] ID:rw_01kz7ssxh8q28hk09y3ycy8kf2 > Verification is your responsibility, not Claude's. Where/why: closes the loop on why bounding tools and subagent scope still requires a human check step. ## Permanent candidates - The deep-read's generated Permanent Candidate for this article ("A development rule belongs in the narrowest mechanism that can reliably enforce its consequence") duplicates an existing vault note almost exactly — see [[3. Resources/Permanent Notes/A Control Is Enforceable Only When the Actor Cannot Reclassify It]], created from a different source (an AI-agent WordPress migration article) and already covering this claim. **Not promoted as a new Permanent note here** — connect instead, per the skill's "check existing Permanents first" rule. ## Processing checklist - [ ] Recheck any command name, precedence rule, or cost figure against current Claude Code docs before using this article as a policy basis - [x] Checked for an existing Permanent note covering the generated candidate claim — found one, linked instead of duplicating - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — a practitioner account of how developer tooling design is adapting to agentic coding workflows - [[3. Resources/Permanent Notes/A Control Is Enforceable Only When the Actor Cannot Reclassify It]] — because evidence-for; this article's CLAUDE.md-vs-hooks distinction is a second, independent instance of the same enforceable-layer principle - Metadata: [[3. Resources/Readwise/Articles/Navigating Claude Code The Full Workflow|Navigating Claude Code - The Full Workflow]]