Skip to main content
Start with this request:
Help me choose where to advertise. I have £10 for an awareness campaign in Great Britain from 1 to 7 October 2026. Show me a plan to review before anything is sent or bought.
This practice task uses fictional sellers held in memory. It does not connect an account, read real inventory, contact sellers, save a campaign, or spend money. The dates and prices are fixed teaching examples, not current offers. It is a candidate exercise awaiting review, not an approved first-value task.

​
Run the exercise

Use a checkout of the repository containing this guide with its workspace dependencies installed. From the repository root, run:
No credentials or model are needed. The command runs a scripted version of the request above and prints a checked JSON receipt. It does not interpret your own text or measure how well a person or model understands the request. The version is seller-discovery@2, with fixture and rubric version 2 and candidate registry version 3. For the application-owned version of the task:
My app will decide what to do. Help it find places to advertise within this budget and date range. Leave approval and next steps to my app.
agency-plan is also available for a small-business planning persona. All three personas exercise the same fixed data; they are three scripted runs, not three people or three independent proofs of product success.

​
Check the plan

The exercise reads both pages of six fictional sellers. Only fictional-newsletter meets the region, currency, budget and readiness rules. The other five are excluded because of price, region, blocked access, missing readiness information, or currency. A listed option is never assumed ready. The draft keeps the £10 limit in GBP, the awareness objective, and the exact flight dates. It requires approval and records zero live actions. Its fixture prices only check the arithmetic; they do not forecast delivery or guarantee inventory. This small teaching object is not a canonical campaign brief or an API request you can send. The deterministic check verifies the complete shortlist, exclusion reasons, page coverage, dates, currency, objective, budget, and approval boundary. The plan and seller data stay in memory and are not copied into the receipt.

​
Recover from the expected error

The runner deliberately tries a negative budget. The fixture must reject it with BUDGET_INVALID. It then restores the original £10 budget and checks the whole plan again. Silently accepting the invalid budget fails the exercise. A real planning workflow should ask you to correct the budget, not increase it without your decision. A plan with no matches keeps that result visible.

​
Read the receipt

Success has event: "candidate_task_completed", job.outcome: "succeeded", and result.scores.fixtureChecks showing three passed checks out of three: execution, correctness and recovery. A failed check produces candidate_task_failed and exit code 1. Invalid command arguments produce a bounded error without echoing the argument. approval and buyerDomainApproval remain null, and kpiEligible remains false. Install, OAuth, intended-account verification and approved first value remain unavailable. activation.receipt and activation.completeThrough are null: no authenticated completion event or complete event stream was observed. Exit code 0 proves only that this local exercise passed. You can save standard output as a receipt. Do not combine it with terminal logs, prompts, credentials or other data. The receipt contains fixed labels, versions, a random receipt ID, times and check outcomes. Reuse that same saved receipt when replaying the observation; a new run creates a new ID.

​
Move to a connected workflow

The local exercise cannot establish account eligibility or approval. A connected run needs an explicitly eligible account, real OAuth and intended-account verification, reviewed seller-discovery and planning contracts, and approval of the exact task and fixture. Until those exist, keep those stages unavailable. For current connected behaviour, use Build an agent and Buyer workflows. Seller discovery is a read; requesting proposals, saving a campaign and launching it are separate actions outside this exercise.