14. 14. Verifying Claims from the Original Thread — Claude Code Guide

This chapter is a line-by-line breakdown of those ~20 points that started the guide preparation. Each statement has been verified against official documentation (docs.claude.com, code.claude.com/docs), the Claude Code changelog, and through Context7 (/anthropics/claude-code/v2.1.89). The goal is to separate facts from speculation so you can cite this thread without worrying about “what if there’s misinformation.”

Legend:

✅ — confirmed by documentation.
🟡 — partially true / has nuances.
❌ — inaccurate or incorrect.
🧪 — relates to experimental feature (behavior may change).

14.1. Summary table

#	Thread claim	Status	Reality
1	Claude Code is a harness around an LLM, not the model itself	✅	See 01., 08.. The harness manages context, tools, hooks, permissions.
2	An LLM by itself is a chatbot; the loop tool_use → tool_result makes it an agent	✅	This is the basic Anthropic API protocol: `stop_reason: "tool_use"` → harness executes → `tool_result` → next iteration.
3	Context is limited to 200k tokens, but can be expanded to 1M	🟡	True, but 1M is only available via alias `opus[1m]` / `sonnet[1m]` and only on Max/Team/Enterprise. See 11.6.
4	Enabling 1M context “by default” is a bad idea	✅	More expensive + empirically quality drops after 300-400k. Anthropic doesn’t recommend “always 1M”.
5	Prompt cache gives huge savings	✅	Cache read = 0.1× input price (10× cheaper). On a 30-turn session, savings are 65-80%. See 02., 11.5.
6	Cache TTL is 5 minutes by default, can be extended to 1 hour	✅	5min cache write = 1.25× input, 1h cache write = 2× input.
7	There’s an “Advisor mode” for planning on Opus	❌	The correct name is `opusplan`. See 02.6., 11.4.2.. After ExitPlanMode, automatically switches to Sonnet.
8	`opusplan` supports 1M window	❌	The plan phase of `opusplan` works in standard 200k, even if 1M is enabled globally.
9	There’s an env variable for autocompact threshold	🟡	Yes, but it’s called `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`, not `CLAUDE_CODE_AUTO_COMPACT_THRESHOLD` as in the thread. Value is a percentage (e.g., `85`).
10	CLAUDE.md has a hierarchy: project, user, local	🟡	Actually 5 levels: managed (enterprise) > user (`~/.claude/CLAUDE.md`) > project (`./CLAUDE.md`) > project local (`./CLAUDE.local.md`, deprecated) > subdirectory CLAUDE.md. See 03..
11	CLAUDE.md supports @-imports	✅	Recursively up to 5 levels. Use `@./path/to/file.md`.
12	Skill is just a markdown file	❌	Skill is a directory with `SKILL.md` + optionally `scripts/`, `references/`, `templates/`. See 04..
13	Hooks are shell commands on lifecycle events	✅	Confirmed. But there are more than 6 events (as in the thread) — actually 28+. See 05.1.
14	A hook can block a model action	✅	Exit code `2` blocks with stderr → returned to model as error. JSON response `{"action": "block"}` — same thing.
15	MCP servers support stdio, SSE, HTTP	✅	All three transports. See 06.2.. In practice, most servers use stdio.
16	Plugins allow packaging MCP + skills + hooks	✅	Manifest `.claude-plugin/plugin.json`, can be published via marketplaces. See 07..
17	Subagents have isolated context and return only summary	✅	Each subagent is a separate API call with its own prefix. Final text is returned to main context. See 09..
18	There are built-in subagents: explore, plan, general-purpose	✅	Full list: `Explore` (Haiku), `Plan`, `general-purpose`, `statusline-setup`, `Claude Code Guide` (if plugin installed).
19	Subagents can run in parallel and in worktree isolation	✅	Parameter `isolation: "worktree"` creates a temporary git worktree. Parallel execution — multiple Agent tool_use in one assistant message.
20	Agent Teams is a full-featured capability for parallel work of multiple Claude	🟡🧪	Real feature since v2.1.32. But experimental, requires `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`. See 10..
21	Tool `Task` is running a subagent	🟡	Was, but in v2.1.63 renamed to `Agent`. `Task(...)` still works as an alias.
22	Opus is the best model for agentic coding	✅	From docs: “most capable for complex reasoning and agentic coding”. Current — Opus 4.7 (released April 16, 2026).
23	Opus 4.7 has a new tokenizer, consumes more tokens	✅	Up to +35% on the same text compared to Opus 4.6. If you had old budget estimates — recalculate.
24	Sonnet is the best price/quality ratio	✅	“best combination of speed and intelligence”. Suitable for 80% of tasks.
25	On Bedrock/Vertex/Foundry, default aliases are shifted back one version	✅	`opus` there → 4.6, `sonnet` → 4.5. Specify full name for latest.
26	Cache doesn’t work if prefix changes (e.g., CLAUDE.md has {{date}})	✅	Cache invalidation triggers on any byte change. So keep dynamics outside the cached part.
27	`gh pr create` is available to Claude via bash	✅	Provided `gh` is installed and authenticated. It’s just a bash command.
28	Permissions have 3 levels: allow / ask / deny	✅	`deny` is final, can’t be bypassed even in `bypassPermissions`. See 08.5.
29	Skills can run from subagents and inherit or fork context	✅	Frontmatter `context: inherit` (sees parent context) or `context: fork` (new context, isolated).
30	”With new models, you need to review CLAUDE.md and skills”	🟡	Not a docs quote, but sound practice. See 11.8..

14.2. Detailed comments on inaccuracies

14.2.1. “Advisor mode” vs `opusplan`

The thread mentioned “Advisor mode” — there’s no such name in the docs. The real command is:

/model opusplan

or right at the start of the session:

claude --model opusplan

📘 Behavior: Opus model in plan mode (read-only tools), after ExitPlanMode automatically switches to Sonnet for implementation. This is the “think with Opus, do with Sonnet” pattern.

⚠️ Known nuance: the plan phase of opusplan does not use the 1M window, even if it’s enabled for the main session.

14.2.2. Env variable for autocompact

Was: CLAUDE_CODE_AUTO_COMPACT_THRESHOLD. Actually: CLAUDE_AUTOCOMPACT_PCT_OVERRIDE.

Takes a number from 0 to 100 (percentage of window fill at which autocompact triggers).

export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=85

Default is around 92%. Can be disabled via 100 (never compress) or via --no-auto-compact flag on startup.

14.2.3. Skill = directory

Yes, a minimal skill is SKILL.md with frontmatter and markdown body. But the structure can be much richer:

.claude/skills/prepare-pr/
├── SKILL.md                  # обязательно
├── scripts/
│   └── lint-and-test.sh      # опционально
├── references/
│   └── pr-template.md        # опционально
└── templates/
    └── changelog-entry.tmpl  # опционально

A skill can call its own scripts and read its own references from its SKILL.md. This turns a skill into a full-fledged “module”, not just a prompt.

14.2.4. List of hook events

The thread had ~6 events. Actually on v2.1.89 there are 28+. Full list in 05.1.. Especially useful ones that people often don’t know about:

SessionStart — for welcome message / pre-flight checks.
Stop — for desktop notifications “model finished, check it”.
TeammateIdle — for Agent Teams.
Notification — customize notifications.
SubagentStop — when a subagent completes.
PreCompact / PostCompact — for logging or customizing compression.

14.2.5. Opus 4.6 → 4.7

At the time the thread was written, the latest version might have been Opus 4.6. Since April 16, 2026, the current version is Opus 4.7:

Same pricing ( $5/$ 25 input/output per MTok).
New tokenizer (+ up to 35% tokens on the same texts).
Improvement on agentic benchmarks, but expect regressions too — review your evals.

Alias opus → 4.7 on API. On Bedrock/Vertex/Foundry — still 4.6 for now.

14.2.6. Tool `Task` vs `Agent`

In subagent call code you may encounter both variants:

Task(subagent_type="Explore", prompt=...)   # старый
Agent(subagent_type="Explore", prompt=...)  # новый, с v2.1.63

They’re equivalent. Inside the harness — it’s the same tool. In new projects, use Agent for compatibility with future versions.

14.2.7. CLAUDE.md hierarchy

The thread mentioned 3 levels. Actually — 5:

Subdirectory CLAUDE.md files are loaded automatically when the model works with files in that subdirectory. This is convenient for monorepos.

⚠️ CLAUDE.local.md is formally deprecated in favor of gitignored CLAUDE.md or just using ~/.claude/CLAUDE.md.

14.3. What was completely correct in the thread

To avoid the impression that “everything was wrong” — no, the thread overall described the system pretty well. Here’s what was definitely accurate:

✅ Main idea of harness vs model. ✅ The tool_use / tool_result cycle as the heart of an agent. ✅ Prompt cache as key savings. ✅ Subagents for browse-heavy tasks. ✅ MCP as the extension standard. ✅ Permissions with allow/ask/deny. ✅ Worktree isolation for safe parallel tasks. ✅ The idea “not always Opus — Sonnet 80%”.

Where the thread author was especially right — in general recommendations. Where they stumbled — in command names and exact env variables. This is typical for Twitter format: the principle sticks, but specifics are easy to forget.

14.4. Verification sources

Each claim in this table has been verified against one or more sources:

📘 Official documentation:

https://docs.claude.com/en/docs/claude-code/overview — public Claude Code docs.
https://code.claude.com/docs/ru/overview — Russian version (synchronized, but sometimes lags).
https://docs.claude.com/en/docs/build-with-claude/prompt-caching — cache details.
https://docs.claude.com/en/docs/agents-and-tools/tool-use/overview — tool use in API.
https://docs.claude.com/en/api/messages — API reference.

📘 Changelog:

https://docs.claude.com/en/docs/claude-code/changelog — for dates and renames (Task → Agent, env variables).

📘 Context7:

/anthropics/claude-code (v2.1.89) — for code examples, current CLI, exact tool names.
/anthropics/anthropic-sdk-typescript — for SDK code in 06. and 12..

📘 Anthropic blog:

Model announcements (Opus 4.7, Sonnet 4.6, Haiku 4.5) — for dates and pricing.

⚠️ In case of source conflict, priority: changelog > docs > blog > Context7. Context7 is good for code, but sometimes lags behind fresh API changes.

14.5. Conclusion: how to read Twitter threads about AI tools

📝 A few heuristics that will save you time:

Always cross-check command names / env variables with docs. They’re easiest to forget, and they change between versions.
Always cross-check numbers. Price, context size, TTL frequency — this is specifics, and it changes too.
Principles are usually correct. “Cache saves”, “subagent isolates context”, “opus is expensive” — these are stable properties, and they don’t break from version to version.
Be skeptical of promotional posts. If the author is selling a course / service on the topic, there’s always exaggeration. Especially suspicious are numbers like “10x productivity” without methodology.
/release-notes is your friend. Every major version (2.1.50 → 2.1.89) brings changes. Reading once a month is too little, twice a month is normal.

14.6. What to do with this guide

This guide is a snapshot as of April 23, 2026, Claude Code v2.1.89. Some things will become outdated in a month, some in a quarter. Use it as:

✅ Reference for concepts (they change slowly).
✅ Template for CLAUDE.md / skills / hooks (adapt to your project).
✅ Checklist of antipatterns.
🟡 Source of specific numbers and commands — but always cross-check with current docs before serious use.

💡 If you find an error — that’s normal, the tooling evolves fast. Best practice: keep your own “corrections map” nearby, and sync with current docs once a quarter.

End of guide.

🚀 Good luck with Claude Code. If this guide helped you avoid even one mistake — it paid for itself.

Back → README (table of contents)

14.1. Summary table#

14.2. Detailed comments on inaccuracies#

14.2.1. “Advisor mode” vs opusplan#

14.2.2. Env variable for autocompact#

14.2.3. Skill = directory#

14.2.4. List of hook events#

14.2.5. Opus 4.6 → 4.7#

14.2.6. Tool Task vs Agent#

14.2.7. CLAUDE.md hierarchy#

14.3. What was completely correct in the thread#

14.4. Verification sources#

14.5. Conclusion: how to read Twitter threads about AI tools#

14.6. What to do with this guide#