Chat Handoff Skill

Generate a self-contained markdown file that lets any AI (Claude or otherwise) continue the current session seamlessly — as if no chat switch happened.

Core Principle

Extract everything from the conversation. Ask the user nothing unless they explicitly give you extra instructions alongside the handoff request (e.g. "make a handoff, also note that the exam date changed to June 20"). If they say something like that, incorporate it. Otherwise, read the conversation and infer everything yourself.

Step 1 — Read the Conversation Thoroughly

Before writing anything, scan the entire conversation for:

• What is this session about? (studying, project, research, work task, planning, etc.)
• Who is the user? Any identity context mentioned (name, role, school, company, tools they use, background)
• What has been completed? Every topic covered, problem solved, task done, decision made
• What is pending? Everything mentioned but not yet done
• What schedule or deadlines exist? Dates, exam days, blocked days, milestones
• What style preferences did the user express? Teaching style, communication preferences, rules for the AI to follow, things they explicitly asked for or pushed back on
• What was flagged for later? Problems to revisit, things the user said were tricky, open questions
• What is the very next thing to do? The exact point where this chat ended

Step 2 — Write the Handoff File

Output a single markdown file to /mnt/user-data/outputs/handoff.md.

Structure it as a prompt for the next AI agent — written in second person ("You are talking to..."), as if briefing a new agent who has never seen this conversation.

Required Sections (adapt names/content to context):

1. WHO YOU ARE TALKING TO

One paragraph. Name, role/background, relevant tools or experience, communication style (e.g. "direct, no flattery, wants blunt feedback").

2. WHAT THIS SESSION IS ABOUT

One paragraph. What are they working on? What's the goal? What event/deadline is driving this (exam, launch, submission, etc.)?

3. CONTEXT & BACKGROUND

Any domain-specific details the next AI needs to know upfront. For example:

• For a placement prep: the company, role, interview rounds, platform, constraints
• For a coding project: the stack, architecture decisions made, constraints
• For research: the topic, angle, what sources have been used
• For work: the task, stakeholders, decisions already made

Keep it factual and dense — no fluff.

4. FULL PROGRESS STATUS

A complete, scannable status table or list. Every item that was planned — mark it ✓ Done or ⬜ Pending. Be exhaustive. Group logically (by topic, module, feature, etc.).

Include sub-details where relevant (e.g. which approach was used, what constraint applies, if it was flagged for revision).

5. FLAGGED ITEMS

Anything the user explicitly said was tricky, confusing, needs revisiting, or should be noted for later. Even one item deserves its own section.

6. SCHEDULE / DEADLINES

If any dates, timelines, or blocked periods were mentioned — list them clearly. Include:

• The deadline or event date
• Blocked/unavailable days
• Remaining available days
• Day-by-day plan if one exists

7. WHERE TO PICK UP

One clear, unambiguous statement: what is the very next thing to do? Be specific — not "continue studying" but "Start with CSS Animations, then move to LC 658."

Also include what to say as the opening line of the new chat (e.g. "Say: 'Hey [name], picking up right where we left off. Today is [date] — let's get into it.'")

8. STYLE / BEHAVIOUR INSTRUCTIONS

Any rules the user set for how the AI should behave. This is critical — if the user pushed back on something, corrected the AI, or stated a preference, capture it here so the next AI doesn't repeat the mistake.

Examples:

• "When the user challenges your reasoning, verify before pushing back — they have been right multiple times"
• "Don't give unsolicited encouragement or flattery"
• "Teach 4-5 topics at a time, then give a copy-pasteable example"
• "One topic at a time. Don't dump everything at once"
• "If code is correct, confirm and move on — don't over-explain"

9. HARD CONSTRAINTS (if applicable)

Domain-specific rules that must always be followed. E.g.:

• No built-in data structures
• Vanilla HTML/CSS/JS only
• No external libraries
• Specific file structure requirements

Only include this section if hard constraints exist.

Step 3 — Filename and Delivery

Save to: /mnt/user-data/outputs/handoff.md

If a handoff file already exists from this session (user is doing a second switch), name it handoff_v2.md, handoff_v3.md, etc.

Then call present_files with the path so the user can download it immediately.

Writing Rules

• Write for any AI reader, not just Claude. Use plain language, no Claude-specific assumptions.
• Be dense and specific, not generic. "CSS Grid — all 12 topics done" is good. "CSS mostly done" is useless.
• Never summarize vaguely. If something was covered, say exactly what was covered.
• Tense: Write the status sections in present tense ("X is done", "Y is pending").
• Opening line of new chat: Always include a suggested first message for the new AI to say — this kills the "new chat" feeling immediately.
• Date awareness: Use the actual current date when writing the schedule and "where to pick up" sections.
• Keep the file self-contained — the next AI should need nothing else to continue.

What NOT to Do

• Do not ask the user questions before generating (unless they gave you extra instructions in the same message)
• Do not include meta-commentary like "I've summarized the session below"
• Do not truncate the progress table — every item must appear
• Do not write in first person ("I covered X") — write as a briefing ("X has been covered")
• Do not forget flagged items — these are often the most important things to carry forward