| name | continue |
| description | Emit a paste-ready continuation prompt that lets a fresh Claude Code session pick up exactly where this one left off. Use when context is filling up and the user says "continuation prompt", "continue prompt", "write me a continuation", "carry this over", "hand this off to a fresh session", "I'm running low on context", or invokes `/continue`. NOT the CLI `--continue` flag — this produces a self-contained handoff brief. |
| argument-hint | [focus] — optional: narrow the handoff to one thread (e.g. 'just the ship PR work'); omit to capture the whole session |
| user_invocable | true |
/continue — write a continuation prompt
The operator is running low on context and wants to resume this work in a fresh session. Synthesize the current conversation into a single paste-ready brief that a cold session can act on immediately.
This is invoked when context is tight, so work from what's already in context — don't go exploring. The conversation is the source. One cheap git check to ground the state is fine; a research sweep is not.
What the fresh session already has (so you don't repeat it)
Inherited automatically:
- The repo's
CLAUDE.md
- Auto-memory for the project (
~/.claude/projects/<hash>/memory/)
- Global skills, MCP servers, settings
NOT inherited:
- This conversation — zero memory of what we discussed, read, decided, or tried
- Files read, diffs seen, uncommitted edits, dead-ends hit
So the brief carries only what this conversation added. Don't restate CLAUDE.md, memory, or house conventions — they load on their own. Every file path, decision, constraint, and failed approach that matters has to be in the brief or it's gone.
Ground the state (best-effort, one pass)
If the session is in a git repo, a single quick check sharpens the "State" section — current branch, uncommitted files, last commit or two, and open PR/CI if relevant. Skip silently if not a repo or it'd cost too much. Never let grounding balloon into exploration.
The brief
Emit it as one fenced code block so the operator can copy it clean. Use these headers; skip any section that's genuinely empty rather than padding it.
Continuing from a prior session that ran low on context. You already have CLAUDE.md,
project memory, and skills — below is only what that session added.
## Goal
<the overall objective, 1-2 sentences>
## State
- Done: <concrete outcomes — paths, SHAs, PR#s. No process narration.>
- In flight: <what's mid-stream right now, incl. uncommitted edits to <files>>
- Branch / PR: <branch>, <PR url if any>, CI <state>
## Next
1. <the immediate next action, concrete enough to start cold>
2. <then this>
## Key facts (this session only)
- <decision made + the one-phrase why>
- <constraint or API shape discovered>
- <dead end already tried — don't repeat: ...>
## Pointers
- <file:line worth opening first>
- <command to re-run / how to verify>
Rules
- Honest state. Done means done and verified. If something's untested or half-applied, it goes under "In flight," not "Done." Don't launder hope into completion.
- Tight. Results, paths, PR#s, SHAs — not adjectives. If a section won't fit in a few bullets, the framing is off, not the cap.
- Self-contained. Assume the reader knows nothing about this chat. A path or decision that lives only in our conversation must appear here.
- Don't re-derive the inherited. No CLAUDE.md recap, no memory recap, no house-style boilerplate.
- If a
[focus] arg is given, scope the brief to that thread and say so in the Goal line.
After emitting
One line below the block: offer to save it to docs/prompts/continue-<YYYY-MM-DD>.md (scratch — not committed) if they'd rather keep it than paste it. Default is print-to-chat; only write the file if asked.
