- Home
- Skills
- Agents & Orchestration
- Crew Delegation Brief Writer: Hand-offs Subagents Can Finish
More screenshots
Works with the AI Agents you already use
Crew Delegation Brief Writer: Hand-offs Subagents Can Finish
Handing work to a subagent fails in two quiet places.
Free
Crew Delegation Brief Writer: Hand-offs Subagents Can Finish
Example session with this skill installed
I need to hand a bug to a background coding agent that works
alone in its own checkout and can't ask me questions. Write the brief I
should send it. What I know: repo https://github.com/example-org/invoice-api
(Python 3.12, FastAPI, pytest). EU customers' invoices round VAT wrong: a
line with 19.995 VAT shows 19.99 instead of 20.00. I think the bug is in
app/billing/tax.py, which does int(x * 100) / 100. I tried swapping in
round(x, 2), but tests/test_invoice_totals.py then failed two cases with
half-cent values (Python rounds half to even). We ship through PRs against
main and CI must be green. Don't change the public API response schema. The
agent must not merge anything. Then, when the agent reports back "Done, PR
https://github.com/example-org/invoice-api/pull/88 uses Decimal with
ROUND_HALF_UP, all tests pass locally, CI still running. Should credit notes
use the same rounding?", write the update I should send my lead.
- Read your context and instructions
- Compiled the crew delegation brief
Below are both messages. I haven't opened the repo or PR #88, so everything about the code and the fix comes from what you told me or what the agent reported.
1. Brief for the background agent
Origin: requested by <your name> via the main session.
REQUEST ID: fix-eu-vat-rounding-20261107
(Reuse this id if this brief is re-sent; it is the same task.)
PROJECT
- Repo: https://github.com/example-org/invoice-api
- Start from: main
- Stack: Python 3.12, FastAPI, pytest
- Tests: run the full suite with `pytest`. tests/test_invoice_totals.py is the main file for this bug.
GOAL
EU customers' invoices must round VAT correctly. Today a line with
19.995 VAT shows 19.99; it must show 20.00. Half-cent values round up,
not down and not to even.
DONE WHEN (all must be true)
1. A line with VAT 19.995 comes out as 20.00 on the invoice. There is a
new pytest test for exactly this case.
2. Every existing case in tests/test_invoice_totals.py passes, including
the half-cent cases, with their current expected values unchanged.
3. The full `pytest` suite passes locally.
4. A PR against main is open and CI on it is green. If CI is still
running when you report, say so. Do not report it as green.
5. The public API response schema is unchanged: same fields, same
types, same serialization format.
CONSTRAINTS
- Do NOT change the public API response schema. If a correct fix seems
to require that, stop and report needs-decision.
- Do not change the expected values in existing tests to make them pass.
If you think an existing expectation is wrong, leave it alone and
explain why in your report.
- No new third-party dependencies. The stdlib `decimal` module is fine.
- Out of scope: credit notes, refunds, non-EU tax paths, and any
refactor of the invoice model. If your change to shared rounding code
would also change those, list exactly what changes in your report.
Don't widen or narrow the change on your own.
STARTING POINTS
- Suspected location (not confirmed): app/billing/tax.py, which does
`int(x * 100) / 100`. That truncates instead of rounding. Check
whether other call sites use the same pattern.
- Known trap: 19.995 cannot be stored exactly as a float (it is really
19.99499999...). So `Decimal(19.995)` built from a float still rounds
to 19.99. Build the Decimal from a string, or keep amounts as Decimal
from where they come in. Your test must prove the 19.995 case really
produces 20.00.
ALREADY TRIED
- Replaced the truncation with `round(x, 2)`. Two half-cent cases in
tests/test_invoice_totals.py then failed, because Python's round()
rounds half to even. Don't use round() for this.
SHIP MODE
One PR against main, from a new branch. Explain the rounding rule and
the float trap in the PR description.
AUTHORITY
- You may: create a branch, change code and tests, run tests, push your
branch, open the PR.
- You may NOT: merge (under any circumstances), push to main, force-push
any shared branch, delete branches, change CI config, or comment on or
modify other PRs or issues.
REPORT BACK
- Outcome word: done / blocked / needs-decision / failed
- PR link exactly as printed
- Commands you ran and their results (pytest summary line)
- CI status at report time: green, red, or still running
- The test you added for 19.995 → 20.00
- Every call site you changed, and whether credit notes or other paths
are affected
- What you did not do or could not check
- Any open question, with the options you see
2. Update for your lead (after the agent's report)
Not done yet: PR open for the EU VAT rounding bug (19.995 showed 19.99, should be 20.00); CI still running, not merged.
- Checked by me: nothing yet. I haven't reviewed the PR or the CI result.
- Reported by the agent (not checked): the fix uses Decimal with
ROUND_HALF_UP, and all tests pass locally.
- Not done / gaps: CI not finished. The agent didn't say whether it
added the 19.995 → 20.00 regression test, whether it avoids the
float-to-Decimal trap, or whether the response schema is unchanged.
I'll check all three before review.
- Needs your call: should credit notes use the same rounding?
Option A: yes, same half-up rule (consistent; note that Decimal's
ROUND_HALF_UP rounds negatives away from zero, so -19.995 → -20.00).
Option B: no, leave credit notes as they are and handle them in a
separate change.
Link: https://github.com/example-org/invoice-api/pull/88
Before you send the update, open PR #88 and check three things: whether there's a test for exactly 19.995 → 20.00, whether the Decimal is built from strings and not floats, and whether the change touches the credit-note code. If those check out and CI turns green, move them to "Checked" and change the first line to "Done, pending review."
On credit notes, I lean towards Option B (a separate change) because it keeps PR #88 small, unless credit notes already go through the same tax.py function. In that case the PR has already changed them, and your lead should hear that. The agent's question suggests it doesn't know which is true, so ask it before you send.
Connects securely to your tools. The creator never sees your data.
About this skill
Handing work to a subagent fails in two quiet places. The brief leaves things out, so the worker guesses: wrong branch, a retry added instead of a fix, a merge nobody approved. Then the report comes back and gets "upgraded" on its way to you: "passes locally" turns into "tests pass", "PR opened" into "fixed", and the worker's guess about the cause into a fact.
Crew Delegation Brief Writer fixes both ends.
Part 1: the brief. Every hand-off gets ten fields in a fixed order: Request id, Project, Goal, Done when, Constraints, Starting points, Already tried, Ship mode, Authority and Report back. Acceptance criteria must be checkable by the worker itself. Suspicions are labelled as suspicions. Error text is pasted verbatim. Irreversible and outward-facing actions (merging, force pushes, deleting, sending, publishing, spending) are named explicitly as allowed or not. The request id is made once and reused on every retry, so a re-send is recognised as the same job instead of starting a second one. A five-question cold-read check catches the gaps before you send.
Part 2: the relay. When the worker reports back, the update opens with the outcome (bad news first), separates what you checked from what the worker only claimed, copies links exactly as printed, lists what was not done, and hands any open choice back to you with its options instead of deciding for you.
It triggers on its own when you ask an agent to delegate, write a task for a background agent, or summarise a worker's result. It scales to the job: a quick lookup needs three fields, a multi-hour change gets all ten.
The package is one SKILL.md plus four worked examples: a weak brief rewritten, a report-only investigation, a retry that keeps its id, and a relay written from a mixed report. No scripts, no network access, no setup, and no dependency on any particular orchestration tool.
Requires: any agent that loads SKILL.md skills, such as Claude Code or another Agent Skills compatible agent. It works with built-in subagents, background or cloud agents, task queues, or a note you paste to a person. Nothing to install and no API keys.
Known limitations
- It writes the brief and the relay. It does not send anything, start a worker or check a PR by itself; your agent's own tools do that.
- A brief can only hold what you or the agent knows. Missing facts are flagged as unknowns, not invented.
- "Checked" in a relay means checked in this session. If the agent has no access to the repo or CI, everything stays under "reported".
- Your own instructions win. If you want a one-line hand-off, you get one.
FAQ
Do I need a particular multi-agent tool? No. The brief is plain text that works for any worker that reads instructions, including a human.
Is the request id really needed? It costs one line and stops duplicate work when a send times out and you retry. Queues that dedupe by id drop the copy; workers that see the same id know it is the same job.
Will it slow small hand-offs down? No. Small jobs get a short brief: Goal, Done when and Report back.
Is it free? Yes, free with no upsell inside the skill.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
It writes the brief and the relay. It does not send anything, start a worker or check a PR by itself; your agent's own tools do that. A brief can only hold what you or the agent knows. Missing facts are flagged as unknowns, not invented. "Checked" in a relay means checked in this session. If the agent has no access to the repo or CI, everything stays under "reported". Your own instructions win. If you want a one-line hand-off, you get one.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free to use
No account required to browse
Trust & safety
Security scan
Security scan passed 10 days ago
- Free with an Agensi account
- Software
- Playwright
Node.js