Back to challenges

Challenge 01

Intelligent
Prompt
Orchestrator

Build a persistent, asynchronous system for chaining LLM work without making the user become the memory, monitor, and debugger.

The problem

You open your editor, write a prompt, and wait. The model responds. You write another. This works fine until the task you are actually trying to do cannot fit inside a single exchange.

The moment your work becomes a chain, the model stops being your tool and you become its memory. You copy outputs into the next input, watch sessions so they do not expire, and discover too late when step three quietly produced bad output.

Now imagine that chain has forty steps, touches twenty files, or needs to keep running while you are asleep.

Queuing prompts is a solved problem. What is not solved is keeping them coherent.

Focus 01

Context chaining

Pass the real, stored output of an upstream task into downstream prompts, then validate it before anything depends on it.

Focus 02

Smart failure recovery

Retry with the specific validation or execution failure included as diagnostic context, instead of replaying the original prompt.

Focus 03

Informed absence

Let users leave the run unattended and return to structured results, blocked tasks, retry history, and a clear failure account.

What you are building

A prompt queue that knows what actually happened.

Build a persistent, asynchronous system that lets a user define a sequence of LLM tasks, submit them, close their laptop, and return to either a clean structured result or an honest, specific account of exactly what failed and why.

A flat list of independent prompts firing in sequence is not enough. Your orchestrator must pass actual outputs forward, block downstream work when an upstream task fails, validate outputs against task schemas, and surface retry attempts with useful diagnostic context.

The hard questions are whether the system passes the actual output of one task into the next, notices when something went wrong before the user does, and reports clearly what happened while the user was away.

Provided resources

Starter material

These files define the task formats and test chains teams should use to prove their orchestrator handles real chained work, not just ordered prompt dispatch.

Markdown

Download

Example task chains

Five ready-to-run chains for testing chaining, validation, blocking, and retry behavior.

YAML

Download

Reference task schema

Canonical YAML format your orchestrator must parse for queue submissions.