| name | session-handover |
| description | Produces a structured 5-part handover document at the end of a complex session, ensuring continuity between AI instances. Triggers on "handover", "session end", "wrap up", or after 10+ substantive exchanges. Outputs an operational briefing the next AI instance can parse in 30 seconds — not a conversation summary.
|
Skill: Session Handover
Version: 1.0
Part of: never-start-over
Companion skills: retex (error post-mortem), todo-tracker (task extraction)
1. Mission
Produce a continuity document so the next AI instance can resume work with zero context loss, no re-debating of closed decisions, and no repeating mistakes already fixed.
A handover is NOT a conversation summary. Summaries are for humans who were absent. Handovers are operational briefings for AI instances that need to execute.
The difference: a summary says "we discussed authentication options." A handover says "D1: JWT over session cookies — stateless, scales horizontally. Rejected: OAuth-only (too complex for MVP timeline)."
2. The 5 Mandatory Sections
Section 1: STATE
Current project situation. Factual, scannable, present tense.
Answer: "Where are we right now?"
Include:
- What works
- What remains to be done
- Current version of deliverables
- Blockers (if any)
Format: short paragraph or factual bullets. No narrative. No "we then proceeded to..."
Section 2: DECISIONS (closed)
Choices made during the session + why. The goal: prevent the next instance from re-opening debates that are already settled.
| # | Decision | Rationale |
|---|
| D1 | ... | ... |
Also include rejected alternatives if the debate was significant (1 line per rejected option + rejection reason). This is critical — without it, the next instance will propose the same rejected approaches.
If the session was purely execution with no architectural or strategic decisions: "No new decisions this session — previous decisions still apply."
Section 3: PITFALLS (this project ONLY)
AI errors encountered during THIS session on THIS project. No generalities, no universal lessons — those belong in a RETEX.
| Context | AI Error | User Fix | Impact |
|---|
| ... | ... | ... | ... |
If no significant errors occurred: "No notable pitfalls this session."
Section 4: DELIVERABLES
Every file produced. Exact name, version, path.
| File | Version | Path / Description |
|---|
| ... | ... | ... |
Include intermediate files if useful for the next session (V0 snapshots, configs).
Mark obsolete files with strikethrough.
Verify version coherence — no jump from V1 to V3 without V2.
Section 5: NEXT
Next concrete action. One line if possible.
If multiple actions: order by priority.
3. Loading a Handover (for the next AI instance)
When a handover document is provided at the start of a session:
- Read the entire document before responding to anything
- Treat DECISIONS as closed — do not re-debate them unless the user explicitly reopens
- Check PITFALLS — if a similar task arises, actively avoid the documented errors
- Reference DELIVERABLES — know what files exist, their versions, their locations
- Start from NEXT — this is your entry point, not a suggestion
If anything in the handover contradicts new user instructions, flag the conflict. Don't silently override.
4. Production Rules
Full scan: Re-read the ENTIRE conversation. Critical decisions often happen in the first 20% of a session — skipping early messages is the #1 cause of incomplete handovers.
30-second rule: The handover must be parseable in 30 seconds by an AI instance discovering the project cold. If you wrote prose, you failed.
Numbering: Use decision IDs (D1, D2...). These propagate across sessions — if the previous handover ended at D4, this one starts at D5. If this is the first session or no previous handover exists, start at D1.
Tone: Factual. Direct. Operational. Write like a military briefing, not a meeting recap. The reader is a machine, not a human who needs emotional context.
5. Template
Use the template in assets/TEMPLATE_HANDOVER.md. Read it BEFORE writing. Do not improvise the structure.
If the template file is unavailable, follow the 5-section structure from Section 2 exactly.
6. Naming Convention
HANDOVER_{{Project}}_{{dd-mm-yy}}.md
7. Metadata (mandatory end block)
---
uri: {{conversation link or ID}}
date: {{dd-mm-yy}}
project: {{name}}
type: HANDOVER
reorientations: {{if AI error caused direction change, note here}}
by: AI + User
---
8. Workflow
When the user triggers "handover":
- Re-read the entire conversation
- Read
assets/TEMPLATE_HANDOVER.md
- Fill in all 5 sections
- Add metadata block
- Create file with correct naming
- Deliver the file
- List session files: OBSOLETE or VALID
9. Anti-Patterns
| Do NOT | Do |
|---|
| Write a chronological summary | Capture current state + closed decisions |
| Copy-paste messages | Synthesize each point in 1 line |
| List generic AI lessons | Document errors specific to THIS project |
| Forget intermediate files | List EVERYTHING produced |
| Write paragraphs of prose | Keep it scannable in 30 seconds |
| Omit the "why" behind decisions | Every D needs a rationale |
| Start decision numbering from D1 every time | Continue from the previous handover |
| Summarize the user's emotions or reactions | Stick to facts and deliverables |
10. Quality Criteria
A handover is good when it passes these 3 tests:
Content test: Could a new AI instance resume the project from this document alone, without reading the conversation? If not, something is missing.
Scannability test: Can you identify the current state, the last decision, and the next action within 30 seconds of reading? If not, it's too verbose.
Continuity test: Does the handover prevent the next instance from (a) re-debating a closed decision, (b) repeating a documented mistake, (c) producing a file that already exists? If any of these could still happen, the handover failed.
A quick self-check before delivering:
Why This Skill Exists
Every AI session produces knowledge that vanishes when the context window resets. Decisions get re-debated. Mistakes get repeated. Files get lost. Version numbers get confused.
This skill was born from 500+ hours of real-world AI collaboration and refined through dozens of iterations. The 5-section format survived because anything less loses critical information, and anything more doesn't get read.