Agent-Skills.md

Agent Skills: Session Handover

Prepare a structured handover document from the current session for a fresh Claude session to continue the work. Use when the user says "/handover", "prepare a handover", "write a handover", "session summary for next time", or when wrapping up a complex debugging/implementation session that will be continued later. Captures problem context, locked decisions, dead ends, running state, key files/commands, artifacts, open questions, and next steps so zero context is lost between sessions.

UncategorizedID: kyzooghost/audit-code-health-skill/handover

Repository

kyzooghost/audit-code-health-skill
kyzooghostLicense: MIT

Install this agent skill to your local

pnpm dlx add-skill https://github.com/kyzooghost/skills/tree/HEAD/skills/handover

Skill Files

Browse the full folder contents for handover.

Download Skill

Loading file tree…

skills/handover/SKILL.md

Skill Metadata

Name
handover
Description
Prepare a structured handover document from the current session for a fresh Claude session to continue the work. Use when the user says "/handover", "prepare a handover", "write a handover", "session summary for next time", or when wrapping up a complex debugging/implementation session that will be continued later. Captures problem context, locked decisions, dead ends, running state, key files/commands, artifacts, open questions, and next steps so zero context is lost between sessions.

Session Handover

Produce a structured handover that gives the next fresh Claude session everything it needs to continue without re-discovery.

Process

  1. Scan the conversation for: problem statement, investigation steps, decisions made, files touched, commands run, current state, failed approaches, and unresolved questions.
  2. Active state capture - Run relevant diagnostic commands (git status, kubectl, build state, etc.) to capture live state. If commands fail or aren't available, fall back to what's known from the conversation.
  3. Write the handover to tmp-docs/handover-YYYY-MM-DD-<topic>.md (use kebab-case topic slug). If user specifies a different path, use that instead.

Output Structure

Every section is mandatory. Write "None" if a section is genuinely empty.

## Session Handover ({date}, session `{first_8_chars_of_session_id}`)

### Problem Statement
# 1-3 sentences. What is being solved and why. Include the user-facing symptom.

### Key Decisions Locked
# Bulleted list of decisions that are FINAL and should NOT be revisited.
# Include rationale in parentheses so the next session understands WHY.

### Dead Ends
# Approaches tried that failed. Format:
# - **What was tried** - why it didn't work
# Prevents the next session from repeating failed approaches.

### Running State
# Table or bullets showing current state of relevant systems/components.
# Note the timestamp/conditions under which state was observed.
# e.g., pod status, deploy state, test results, feature flag state.

### Key Files & Commands
# Table: File | Repo/Location | Why it matters
# Then a fenced code block of the most useful diagnostic/operational commands.
# Include ACTUAL paths, ACTUAL command strings - not placeholders.

### Artifacts
# Table: Link | Type | Status | What it does
# PRs, commits, Jira tickets, Confluence pages created or referenced.

### Open Questions
# Things not yet answered that the next session may need to investigate.
# Include enough context for the question to be actionable standalone.

### Next Steps
# Numbered, ordered list of what to do next.
# Each step must be specific and actionable.
# Include verification criteria: how to confirm the step succeeded.

Guidelines

  • Be concrete, not abstract. Include actual file paths, actual command strings, actual error messages. "Check the logs" is useless; kubectl logs -n cbp-web -l app=commbiz-web --since=5m | grep -i "statusCode" is useful.
  • Capture the "why" behind decisions. The next session has no conversation context - it needs to know WHY an approach was chosen or rejected.
  • Running state is a snapshot. Note when it was captured so the next session knows if it might be stale.
  • Keep it scannable. Tables for structured data. Bullets for lists. No paragraphs where a table row suffices.
  • Commands must be copy-pasteable. No pseudocode or partial commands in the Key Commands section.
  • Dead ends save time. Even a one-liner like "Tried X, failed because Y" prevents hours of repeated investigation.