suggest-split
Skill4 starsChanged 20 days agoMITView on GitHub
Detect when a conversation has drifted across multiple unrelated topics and recommend whether to start a new conversation. Identifies the optimal split point and summarizes what to carry forward. Use when a session has been running long or has covered many different tasks.
In this repository7 other files it gives its agents.
- context-healthSkill
- find-bloatSkill
- history-trimSkill
- inspect-contextSkill
- mcp-auditSkill
- slim-instructionsSkill
- token-forecastSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 43
Write a ready-to-paste context block the user can paste at the start of their new session:
- Very specific: line 65
Follow with one line: Context is coherent.
- Distinctive: line 81
Do NOT recommend a split if:
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Conversation Split Advisor
Analyze the current conversation and determine whether it has drifted enough that starting a new session would meaningfully improve context quality.
Analysis process
- Map the conversation topics — scan the full history and identify distinct task areas or subject shifts. A "topic" is a coherent unit of work (e.g., "debugging the auth middleware", "writing README", "reviewing PR #42").
- Find the current task — what is the user actually working on right now? This is the anchor.
- Score past topics by carry-forward value:
- Essential (carry forward): decisions made, constraints discovered, code written that is still active
- Partial (summarize): context that's partially relevant but could be distilled to 2–3 sentences
- Dead weight (discard): earlier topics completely unrelated to the current task
- Assess the split case — is there a clear turning point where the conversation shifted to the current task? If yes, that's the split candidate.
Output format
If a split is recommended:
✂️ Split recommended — conversation has drifted across {N} topic(s)Then:
Topic map (chronological):
| Turns | Topic | Status |
| 1–8 | Debugging login timeout | 🗑️ Discard — resolved, no longer relevant |
| 9–14 | Reviewing PR #42 | 📋 Summarize — 2 decisions still apply |
| 15–now | Refactoring auth middleware | ✅ This is your current task — carry forward |
What to carry into the new conversation:
Write a ready-to-paste context block the user can paste at the start of their new session:
--- CONTEXT FROM PREVIOUS SESSION ---
[2–5 bullet points covering only the decisions, constraints, and code that are still relevant to the current task. Be specific: file names, function names, exact decisions made.]
--- END CONTEXT ---Split instructions:
1. Copy the context block above
2. Start a new Claude Code session
3. Paste the context block as your first message
4. Continue with: [one sentence describing where to pick up]If no split is needed:
✅ No split needed — conversation is focused on a single task ({N} turns, ~{tokens} tokens used)Follow with one line: Context is coherent. Continue in this session.
If the conversation is too short to evaluate (< 5 turns):
ℹ️ Conversation is too short to evaluate for splitting ({N} turns).Decision thresholds
Recommend a split if ANY of the following are true:
- The conversation has covered 3 or more distinct topics
- More than 40% of tokens come from turns unrelated to the current task
- The current task started after turn 15 and the earlier turns add no carry-forward value
- Context is over 70% full and early turns are clearly dead weight
Do NOT recommend a split if:
- The conversation has stayed on one topic throughout
- Earlier turns contain decisions or code that are still actively referenced
- The session is under 10 turns
$ARGUMENTS
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

