메뉴
Skill for evaluating whether a conversation has accumulated enough irrelevant context to warrant starting a new session. Applies structured criteria to distinguish valuable accumulated context from bloat that degrades response quality.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when the user points at a GitHub repo (an agent skill, MCP server, or "Claude plugin"-ish project) and wants to install or register it. Routes between native plugin install, MCP server registration, and adding to the user's third-party marketplace. Triggers on phrases like "install this from GitHub", "add this skill to my setup", "register this as a third-party plugin", "is this available as a Claude plugin".
Reference and worked example for the CONTEXT.md / CLAUDE.md two-file workflow. Use when the user asks "what is CONTEXT.md", how the human-context vs agent-briefing split works, or wants a copyable example pair to seed a new repo. Pairs with the operational skills (context-to-claude, claude-to-context, add-to-context, remove-from-context, manual-context-update, chunk-repo-context) — this skill explains the model; those skills do the work.
Use when the user wants to prune and chunk a bloated CLAUDE.md into a structured context/ store with pipeline folders and subagent templates. Triggers on phrases like "chunk claude", "prune CLAUDE.md", "set up context store", "split CLAUDE.md".
Use when the user wants to strip AI-generated style tics from documentation — emojis, excessive lists, all-caps filenames, fragmented structure. Triggers on phrases like "declaude the docs", "clean up documentation style", "remove emojis from docs", "apply style guide".
Use when the user wants to install a new user-level slash command at ~/.claude/commands/. Triggers on phrases like "install a slash command", "add a /command", "create a user-level slash command", "save this as a slash command".
Save a Claude Code plan (from ExitPlanMode, design discussion, or multi-step task) to persistent storage. ONLY invoke when the user explicitly asks to save/stash/store a plan (e.g. "save this plan", "stash the plan", "store the plan in the planning repo"). Do NOT auto-invoke after planning. Routes to the active project's `planning/` folder when one exists, else to the user's configured central planning repo.
| name | context-break-evaluator |
| description | Skill for evaluating whether a conversation has accumulated enough irrelevant context to warrant starting a new session. Applies structured criteria to distinguish valuable accumulated context from bloat that degrades response quality. |
This skill teaches you how to assess whether continuing a conversation will produce better or worse results than starting a new one. Context management is a critical but often overlooked factor in getting high-quality responses from AI agents.
Every message added to a conversation increases context length. Longer context is not inherently better. Context becomes harmful when:
Starting a fresh conversation clears this noise. A well-timed context break often produces better results than continuing in a degraded session.
When asked to evaluate whether a user should start fresh, apply these criteria in order:
Has the primary task that opened this conversation been completed? If yes, and the user is moving to a new unrelated task, recommend a new conversation.
Estimate what percentage of the conversation history is directly relevant to the user's next request. If less than 30-40% is relevant, the irrelevant majority is likely to cause drift and degradation.
Count significant failed attempts, retried tool calls, conflicting instructions, or explicitly corrected misunderstandings. More than two or three of these in a single thread is a strong signal to start fresh.
If the conversation has touched more than two or three distinct, unrelated topics, the context is fragmented. Fragmented context confuses future responses.
Very long conversations (subjectively: many dozens of exchanges) accumulate compounding noise. Apply extra scrutiny before recommending continuation.
Continue if: the current task is ongoing, prior context is largely relevant, and there are few accumulated errors.
Start fresh if: the prior task is done, context relevance is low, significant errors have accumulated, or topics have diverged substantially.
When in doubt: recommend a fresh start. The cost of starting fresh is low (a brief re-orientation). The cost of degraded context is subtle but persistent.
When recommending a fresh start, always offer to generate a handoff summary — a compact paragraph the user can paste into the new session. A good handoff summary includes:
Keep handoff summaries under 150 words.
Be direct. Users asking this question want a verdict, not a hedge. Commit to a recommendation and explain it briefly. Avoid phrases like "it depends" without immediately resolving what it depends on.
