Every agent-to-human and agent-to-agent baton pass — context dropped, work duplicated, escalations landing where nobody looks.
Free & open · no signup · read-only — it ends by asking · nothing leaves your machine
You are working inside this repo. Mission: trace every baton pass in the system — agent to human, human to agent, agent to agent — and find where context drops, work duplicates, or the baton lands where nobody is standing.
Read-only pass. Your only write is the report file.
Phase 2 sweeps the codebase through every one of these, citing file and line for each finding.
the receiver missing what the sender knew: a human shown a conclusion without the evidence; agent B without agent A's constraints and failed attempts
escalations posted to channels nobody watches; agents waiting on human input with no timeout, reminder, or fallback
receiver redoing sender's work because the handoff didn't carry it, or trust in it
does the sender know the baton was caught? Fire-and-forget passes that fail silently (see 03)
a human fixes agent output: does the fix flow back into prompts, memory, or evals — or evaporate (see 45)?
pending handoffs: countable, aged, alarmed? Or discovered when a user complains
agent emits JSON where the human needs prose; human writes prose where the agent needs fields; translation done badly or not at all
One structured report at the repo root — or in reports/, if you keep one — the same shape every time, ready for a teammate — or the optional Studio — to act on.
HANDOFFS.md already exists from a previous run, read it first and lead with what changed since.Every Goal Prompt follows the same four steps, so results are consistent and repeatable — no matter which one you run.
Enumerate them.
Context drops — the receiver missing what the sender knew.
Rank by stall cost.
Create HANDOFFS.md at repo root.
Copy it and paste it into your agent inside the repo you want checked.
Install the goal plugin once — two commands — then just type /goal:handoff-audit.
/plugin marketplace add GhostlyGawd/goal-prompts/plugin install goal@goal-promptsOr install only this Goal Prompt as /goal-handoff-audit:
curl -fsSL https://goal-prompts.vercel.app/install | BRIEF=41 shLet an agent fetch it mid-conversation, or pull the raw Goal Prompt by URL.
https://goal-prompts.vercel.app/raw/41.mdNothing hidden — this is the whole Goal Prompt, verbatim. Read it in a minute, edit it, or copy it as-is.
# Goal: Handoff Audit You are working inside this repo. Mission: trace every baton pass in the system — agent to human, human to agent, agent to agent — and find where context drops, work duplicates, or the baton lands where nobody is standing. Read-only pass. Your only write is the report file. ## Phase 1 — Map the handoffs - Enumerate them: escalations, approvals, notifications, task submissions, corrections, sub-agent calls, pipeline stages. - For each: what payload transfers, in what format, through what channel, with what deadline. - Which handoffs are load-bearing — the product stalls if they stall? ## Phase 2 — Audit through 7 lenses 1. **Context drops** — the receiver missing what the sender knew: a human shown a conclusion without the evidence; agent B without agent A's constraints and failed attempts 2. **Dead ends** — escalations posted to channels nobody watches; agents waiting on human input with no timeout, reminder, or fallback 3. **Duplication** — receiver redoing sender's work because the handoff didn't carry it, or trust in it 4. **Acknowledgment** — does the sender know the baton was caught? Fire-and-forget passes that fail silently (see 03) 5. **Correction flow** — a human fixes agent output: does the fix flow back into prompts, memory, or evals — or evaporate (see 45)? 6. **Queue visibility** — pending handoffs: countable, aged, alarmed? Or discovered when a user complains 7. **Format mismatch** — agent emits JSON where the human needs prose; human writes prose where the agent needs fields; translation done badly or not at all ## Phase 3 — Curate - Rank by stall cost: load-bearing handoffs first - Every fix names the mechanism: payload spec, timeout + fallback, ack, digest ## Phase 4 — Report Create `HANDOFFS.md` at repo root: 1. **Handoff map** — from → to · payload · channel · deadline · load-bearing? 2. **Findings** — each: handoff · lens · what's lost · fix 3. **Payload specs** — for the worst 3, the exact fields the receiver needs 4. **Timeout and fallback table** — every human-wait, bounded 5. **The one handoff to fix first** Start the report with today's date. If `HANDOFFS.md` already exists from a previous run, read it first and lead with what changed since. ## Rules - A handoff is a contract: payload, deadline, and acknowledgment, or it's a hope - Escalation paths must terminate at a human who will actually see it - No handoffs between humans, agents, or systems in this repo? Say so in a one-paragraph null report and stop — a null result is a valid finding. - If a `reports/` directory exists at the repo root, write the report there instead of the root. - Before asking, present the top findings as a ranked list in plain words - Report only — end by asking which handoffs to fix
reports/ directory exists at the repo root, write the report there instead of the root.Crash the automation at step 3 of 7 on paper — resumability, idempotency, duplicate side effects, poison inputs, dead letters.
LLM calls that should be plain code, rigid code that should be judgment — redraw the line between deterministic and model-driven.