Nightly Vibe · AI coding control room

Split the task before context rot turns an agent run into review debt.

For solo operators directing coding agents: choose the mission, mark context-rot symptoms, attach the evidence packet, and copy a lane plan that a fresh agent or human can execute without rereading the whole saga.

Name the mission and stop condition

One sentence is enough. The tool turns it into a handoff contract.

Task brief

Mark context-rot symptoms

These are reasons to split/reset instead of asking the same thread to “just finish.”

Add operating constraints

Constraints prevent the next prompt from becoming another unbounded heroic run.

Build the evidence packet

Fresh agents need receipts, not lore. Select only evidence you can actually paste or attach.

Copyable artifact

Agent lane plan and restart prompt

Paste this into a Kanban card, PR handoff, or a fresh agent thread before the next run starts.

Symptoms50

Constraints26

Evidence58

CONTEXT BUDGET SPLITTER
Verdict: SPLIT INTO LANES
Mission: Feature build
Brief: Ship the next user-visible slice without letting the agent bury risk in a giant diff.
Overload: 80/100 · evidence coverage 58/100 · recommended lanes 3
Default lane model: Builder → Reviewer → Finisher
Success line: A user-visible slice ships with tests or a local smoke proof.

Symptoms to repair:
- Review overload: Split cosmetic, plumbing, and behavior changes into separate commits.
- False done: Require command output, screenshot, or fixture before summarizing success.
- Missed defects: Demand a review map: files changed, risk per file, and proof per claim.

Operating rules:
- Cap the next run to one sharp outcome and one verification command.
- Write a handoff packet before doing more implementation.

Evidence packet to include:
- Original ask / acceptance criteria: Paste the exact user requirement and the non-goals.
- Current diff + touched files: Include git status, key files, and what each file is supposed to prove.
- Build/test output: Attach the latest passing/failing command output; mark pre-existing failures.

Split plan:
- Lane A: reproduce/inventory the current state without editing.
- Lane B: implement only the narrowest fix or slice.
- Lane C: review proof, merge risk, and handoff gaps.

Next prompt:
You are taking over this task. First restate the success line, inspect only the evidence packet above, then execute the assigned lane. Stop and report if the lane needs context outside the packet.