Chat snapshot skill

Chat Snapshot Skill

Saves and restores full conversation context as a compact, portable JSON file. Works across Claude, ChatGPT, Gemini, or any LLM.

---

Command: /export

When the user types /export (or /snapshot, /save, /compress):

Step 1 — Warn about cost

Say briefly: "Compressing now — this uses some tokens to read the full chat. Save early and often next time!"

Step 2 — Analyze the full conversation

Read the ENTIRE conversation from the very first message to now. Do not skip anything.

Step 3 — Generate a filename

Derive a short, meaningful filename from the conversation topic. Use the first 4–6 words of the goal, lowercase, hyphenated. Examples:

• "python-finance-tracker-cli"
• "claude-skill-chat-snapshot"
• "react-dashboard-dark-mode"

Never use "chat-snapshot" as the filename. Always reflect the actual topic.

Step 4 — Build the JSON snapshot

Output ONLY this exact JSON structure, filled in accurately. Be ruthless about compression — every field must be concise. No filler, no repetition.

{
  "snapshot_version": "1.0",
  "suggested_filename": "<topic-based-filename>.json",
  "exported_at": "<ISO timestamp if known, else 'unknown'>",
  "llm_source": "Claude",
  "goal": "<1-2 sentence summary of what the user is trying to accomplish overall>",
  "context": "<Key background: who the user is, their stack, constraints, preferences, decisions already made>",
  "progress": [
    "<Completed milestone 1>",
    "<Completed milestone 2>"
  ],
  "artifacts": [
    {
      "name": "<filename or component name>",
      "language": "<language if code>",
      "description": "<what it does>",
      "status": "complete | partial | broken"
    }
  ],
  "open_issues": [
    "<Bug or unresolved question 1>",
    "<Bug or unresolved question 2>"
  ],
  "next_steps": [
    "<Immediate next action>",
    "<Following action>"
  ],
  "key_facts": {
    "<any important technical detail>": "<value>",
    "<language/framework/tool>": "<version or note>"
  },
  "full_summary": "<3-5 sentence narrative of the entire conversation so far, enough for a new LLM to understand the full picture>",
  "resume_prompt": "<See format below>"
}

Resume prompt format

The resume_prompt must be fully self-contained — it should work when pasted into ANY LLM even if the skill is not installed. Use this format, under 200 words:

I'm going to give you my conversation context from a previous chat. Please read it carefully and resume helping me from where we left off. Do not ask me to re-explain anything below.

PROJECT: <1-line description>
STACK: <language, frameworks, tools, versions>
DONE: <bullet list of completed work>
ARTIFACTS: <files/components created and their status>
OPEN ISSUES: <bugs or unresolved questions>
NEXT ACTION: <exactly what to do next, specific>

Please confirm you've read this and tell me what we'll work on first.

Do NOT include any explanation inside the JSON. Output only valid JSON.

Step 5 — Instruct the user to save it

After outputting the JSON, say:

✅ Snapshot ready. Save the JSON above as <suggested_filename> (shown in the suggested_filename field).

To resume in Claude: Start a new chat, upload the file, and say "resume from this snapshot." To resume in ChatGPT / Gemini / any LLM: Copy the resume_prompt field and paste it as your first message. No upload needed.

---

Resuming from a snapshot (Claude only)

When the user uploads a .json file and it looks like a chat snapshot (has fields like goal, progress, next_steps), OR when they say "resume", "restore", "continue from this", "here is my snapshot":

Step 1 — Parse the snapshot

Read the uploaded or pasted JSON.

Step 2 — Restore context silently

Do NOT ask the user to explain anything. Treat the snapshot as full working memory immediately.

Step 3 — Confirm restoration

Reply with:

🔁 Context restored. Goal: <goal> Done: <progress as bullets> Open issues: <open_issues as bullets> Up next: <next_steps[0]>

Ready — what do you want to work on?

---

Incremental Export

If the user has a previous snapshot and wants to update it:

• Ask: "Do you have a previous snapshot? Upload it and I'll only summarize what changed."
• If yes: read the old JSON, identify only what is new or changed since then, merge and output the full updated JSON.
• This saves tokens — no need to re-summarize unchanged content.

---

Tips printed after every /export

Always append this after the snapshot:

💡 Pro tips:

• Export every 20–30 messages — don't wait until quota is almost gone
• In ChatGPT/Gemini: just paste the resume_prompt field — no upload, no skill needed
• In Claude: upload the JSON file and say "resume from this snapshot"
• For updates, keep your last snapshot handy to do an incremental export

---

Reference files

• references/sample-snapshot.json — example of a complete snapshot output
• references/resume-prompt-template.md — guide for writing good resume prompts