context-check

Skill: Check context window

Purpose: proactively assess context and recommend the right action before quality degrades.

Process

Step 1 — Read state

Claude cannot run /context itself because it is a built-in command only the user can invoke. Ask the user:

Need to run /context in the terminal and send back the output (% and breakdown).

After the user sends /context output, analyze it per the steps below.

Step 2 — Analyze by threshold

Source for % thresholds (multi-author, verified):

• <30/<40/60% + "dumb zone": Dex Horthy at MLOps Community (2026-03-24)
• 300-400k tokens context rot (1M model): Thariq Shihipar (Anthropic Claude Code team) via howborisusesclaudecode.com
• 155k tokens auto-compact (200k window): Boris Cherny X tweet
• ~95% auto-compact trigger (official): Claude Code sub-agents docs §Auto-compaction — "auto-compaction triggers at approximately 95% capacity"; lower it with CLAUDE_AUTOCOMPACT_PCT_OVERRIDE.
• Full citations + nuance by task complexity: docs/REFERENCE.md §16. Apart from the ~95% auto-compact trigger, Anthropic does not publish official % thresholds for the zone/quality bands below.

% context	State	Recommended action
<30%	🟢 Aggressive zone	Target for experienced users
30-40%	🟢 Sweet spot	Newcomer target — "shoot to keep it under 40%" (Dex)
40-60%	🟡 "Dumb zone" begins	Performance degrading — plan to wrap up current phase
60-77%	🟠 Wrap up actively	/compact OR /handoff → /clear + new brief
~77%	🔴 Critical zone (Boris claim 155k)	Compact proactively — official auto-compact triggers ~95% (CLAUDE_AUTOCOMPACT_PCT_OVERRIDE to lower), but quality degrades here
>90%	⛔ Hard limit	STOP large task immediately, brief + new session

Step 3 — Analyze by group

/context breaks output down by group (system, memory/CLAUDE.md, skills, MCP tools, conversation, file content). Find any group consuming unusually high context:

High-consuming group	Cause	How to reduce
Memory (CLAUDE.md + rules) >10%	CLAUDE.md / rules too long	Prune them, move less-used sections to REFERENCE.md
MCP tools >15%	Too many MCP servers enabled, not used	claude mcp list then disable those not needed for this session
Skill descriptions >5%	Too many auto-discovered skills	Set disable-model-invocation: true for rarely used skills
Conversation history >40%	Lots of tool output / dead-end paths	/compact now
File content >25%	Too many large files @-referenced	/clear + only reference files that are needed

Step 4 — Recommend action

Give 1 main recommendation with a reason — do NOT list 5 options for user to choose from:

Example output:

Context is at 73%. Conversation history occupies 45% — mostly from long tool output in a previous debug session. Recommendation: run /handoff to capture key decisions in 5 lines, then /compact keep the brief, drop debug logs. Then continue the current task in this session. Estimated context after compact: ~25%.

Choosing /compact vs /clear

When to use /compact	When to use /clear
In the middle of a task, need to keep thread	Finished a task, moving to a completely different one
Important decisions and file paths must survive	History is not needed
Context 40-60% ("dumb zone")	Context >77% (auto-compact zone) or session is muddled
Many dead-end debug paths to clean up	Committed and done, starting a new feature

Golden rule: /compact = compress, /clear = wipe entirely. Confusing /clear with /compact = losing context that must be re-explained. Confusing /compact with /clear = carrying garbage into the new task.

When context is corrupt / Claude is confused

Symptoms:

• Claude repeatedly references old files/decisions.
• Claude forgets rules in CLAUDE.md (e.g., still writing English comments even though Vietnamese was configured).
• Fix attempted twice and still wrong.
• Internal server error / ECONNRESET / "Chat has reached its limit".

→ Do NOT /compact (compacting dirty context = continuing dirty). Must:

1. /handoff --save (or copy-paste brief externally).
2. /clear or exit and open a new session.
3. Inject brief into the first prompt.

Long-term tips

• Set a custom status line showing context %: code.claude.com/docs/en/statusline.
• Audit ~/.claude/CLAUDE.md periodically (monthly): remove lines no longer needed.
• Large projects: use subagents (use a subagent to investigate ...) to keep main context clean.
• Large tool output (build log, JSON dump >5KB): redirect to a file instead of dumping into chat: npm test > /tmp/test.log 2>&1 && tail -50 /tmp/test.log.