Как НИКОГДА не упираться в лимиты | Claude и Codex

Main Topic

A guide on how not to hit context/usage limits in chat-based AI coding assistants (Claude, Claude Code, Codex, Cursor, etc.) by reducing token consumption and optimizing chat/workflows.

---

Key Technological Concepts Explained

Tokens

• Tokens are the basic units of text consumed by the model; every character/punctuation counts.
• Approximate rule of thumb:
  • ~3/4 of an English word per token
• Language note:
  • Russian is less efficient: 1 Russian word ≈ 2–3 tokens
  • So Russian text can cost ~2–3× more tokens.

Context window / memory in a chat

• The model must re-read all prior messages every time it generates a response.
• In long chats, most tokens are spent on re-processing history, not producing the new answer.

Context degradation

• As the chat grows, the model can “forget” or become less sharp.
• The described effect is getting “dull” around ~100 messages.

Token cost asymmetry

• Input tokens are cheaper than output tokens (claim: output can be ~5× more expensive).
• Because the model generates output word-by-word and re-processes prior generated text, long answers cost significantly more tokens.

---

Practical “Life Hacks” and Product Features (How-To)

1. Don’t ask unnecessary questions

   • Avoid “re-asking” after an incorrect answer with extra complaints or redundant rework, which burns tokens faster.

2. Use rollback features in Claude Code / Codex / Cursor

   • If the agent’s edits are a mess, use rollback instead of continuing from a bad state.
   • Rollback modes (conceptually):
     • Rewind conversation: rolls back dialogue history to a point.
     • Rewind code: rolls back code/files to an earlier point while keeping chat history.
     • Full rollback / fork conversation: forks the chat and restores code + conversation from before the mistake.
   • Goal: reduce wasted context and avoid forcing the model to redo everything.

3. Edit earlier messages instead of re-asking

   • Example: change “Tula” → “Irkutsk” via an inline pencil edit so the assistant rewrites without expanding the dialogue.

4. Ask multiple questions at once (when appropriate)

   • Bundling reduces how often the model must re-run long chat history.
   • Caveat: not for very large tasks.

5. Use a “listen/side question” style (non-blocking queries)

   • In Codex-like terminals: start a task, then use a “listen” mode (or similar) to ask a clarifying question without interrupting the main generation thread.
   • Note: after using Escape, both the main and side content are removed from the visible chat (as described).

6. Force short, direct answers via settings

   • Configure Claude/Codex to respond without fluff to reduce output token usage.

7. Create new chats periodically

   • Don’t cram a full multi-stage project into one conversation.
   • Recommended structure: one big task per chat; finish it, then start a new chat for the next phase.
   • Use Markdown instruction files when you need context carried forward.

8. Use subagents for heavy analysis

   • For tasks like an SEO audit:
     • delegate deep analysis to a subagent
     • only the final summary returns to the main chat
   • This is described as available in both cloud and codex environments.
   • Subagents/agents are positioned as instructions defining model, tools, and task scope.

9. Use Skills (Markdown-based instructions)

   • Create reusable “skill” instructions for recurring tasks so the model doesn’t re-derive the procedure each time.
   • Skills/agents only load their title + description until needed, reducing baseline context load.

10. Stay under a token budget target (~120k)

    • Even if the context window is large (e.g., up to ~1M tokens), the practical recommendation is limiting usage to around 120,000 tokens due to ongoing history reprocessing and slowdown/context degradation.
    • An update is mentioned that increased context window size, but the degradation/slowdown issue is argued to remain.

11. Convert documents/files to Markdown

    • Don’t feed PDFs/Word/HTML directly when possible.
    • Reason: PDFs/Word include lots of non-text metadata (font info, coordinates, layout, images, profiles), which becomes extra tokens.
    • Markdown is treated as “native”/preferred and can drastically reduce token footprint (example given: 15M tokens vs 8k tokens).

12. Manage Claude.md / Agents.md size (keep it small)

    • These global/project instruction files are loaded at session start.
    • If too large, they waste tokens immediately and can pollute context.
    • Rule of thumb: keep them to ~200 lines, and move large blocks (e.g., design systems) into separate Markdown files referenced when needed.

13. Disable unnecessary tools (MCPs/connectors/features)

    • Extra tools consume tokens and increase latency simply by being available.
    • Example: disable an MCP like a screenshot/UI-checker if not needed.
    • Described as navigating to MCP list and pressing disable.

14. Pick lightweight models for lightweight tasks

    • Use smaller models for routine edits (example tiers: haiku/sunet/opus).
    • Bigger models cost more tokens/thinking; reserve them for complex work.

15. Turn on planning mode before execution

    • Enable a plan-first workflow: 1) the agent reads inputs/specs 2) generates a plan without changing code 3) you approve/adjust 4) only then it executes
    • Goal: reduce major mistakes and rework, saving tokens.

16. Use a planning-focused plugin (“Superpers”)

    • A plugin (reviewed elsewhere by the speaker) that uses built-in skills for a structured workflow:
      • brainstorming → planning → implementation → testing
    • Recommended mainly for larger tasks; considered overkill for trivial ones.

---

“Stretching” Usage Limits in Claude (5-Hour Window)

Daily 5-hour limits window

• Claude has a daily 5-hour limits window (Pro/Max/Team/Enterprise mentioned).
• The window starts from the first message sent.
• If you hit limits early, the quota resumes after the 5-hour window ends.

Workaround: Claude Routines

• Create a routine that sends a small trigger message (e.g., at 6:00 AM) so the 5-hour window resets and you extend working time with fewer interruptions.

Status update on limits (recent change described)

• On May 6, 2026, Anthropic reportedly expanded infrastructure (Memphis data center access) and doubled the 5-hour limits for Claude products (Pro/Max/Team/Enterprise).
• Also removed peak-hour restrictions and raised rate limits for Opus models, making it harder to hit limits.

---

Speaker’s Auxiliary Resources / Recommendations

• The speaker mentions creating full guides for:
  • Claude Code
  • Codex
• Promotes their Telegram channel for additional tips, including legal/web development info and recurring AI news/podcast releases.
• Mentions they may add missing items to the Telegram channel.

---

Main Speakers / Sources

• Primary speaker: the video’s narrator/author (unnamed in subtitles).
• Referenced sources: unnamed developer(s) who measured token usage in long chats (example claim: ~98.5% tokens spent on processing prior responses).
• Company referenced: Anthropic (Claude).
• Tools/products referenced: Claude, Claude Code, Codex, Cursor, plus MCPs, Agents/Subagents, Skills, Claude Routines.