Challenge 5: One-Shot an Agentic Workflow

Global Hack Week: Agent Week | Backboard.io

Time needed: 15-20 minutes (most of it spent writing one very good prompt)
Difficulty: Medium. The code is not the hard part. The prompt is.
What you need: Everything from Challenges 1 to 4. You should have already built and run a basic agent.

CHALLENGE WALKTHROUGH VIDEO HERE

What does “one-shot” mean?

One shot = one prompt. You write a single prompt, hand it to R-CLI or Backboard Studio, walk away (metaphorically), and it builds the entire working thing with no follow-up instructions from you.

In Challenge 4 you could nudge, correct, and iterate. Not this time. If the result is broken or wrong, you do not patch it with follow-ups. You go back, improve the prompt, and fire again from scratch. Your prompt is your program.

Why practice this? Because it forces you to think like an architect: every requirement, every edge case, every definition of “done” has to be in the prompt before you press Enter. People who get good at this build absurdly fast.

> House rules: if your build agent asks you a clarifying question, answering it is fine. Fixing its output with new instructions is not. Re-rolling with an improved prompt is always allowed, and honestly, expected.

What is an “agentic workflow”?

A workflow is a job with multiple steps that happen in sequence, where each step feeds the next. An agentic workflow means AI is doing the steps, making decisions along the way, not just answering one question.

A chatbot answers. A workflow does: gather → think → produce → deliver.

Your mission: one-shot a workflow that takes an input, runs at least three AI-powered steps, and produces a finished output file. It should run start-to-finish with one command.

Pick a workflow (or invent your own)

Research Reporter: topic in → web-searches for current info → summarizes findings → writes a clean markdown report with sources → saves report.md.
Study Pack Generator: any topic in → explains it at beginner level → generates 10 flashcards → makes a 5-question quiz with answer key → saves it all as files.
Hackathon Idea Machine: a theme in → brainstorms 10 project ideas → scores them on difficulty and wow-factor → picks the best 3 → writes a mini pitch for each.
Code Explainer: a code file in → explains what it does in plain English → flags bugs or risks → suggests improvements → writes EXPLAINED.md.
Daily Brief: a list of interests in → searches the web for fresh news on each → summarizes → compiles a personal morning briefing file.

All of these hit the bar: input, 3+ AI steps, file output.

Anatomy of a one-shot prompt

Before the example, here is the checklist your prompt must cover. Miss one and you will feel it:

The goal in one sentence at the top.
The exact tech: Backboard API via backboard-sdk, key from BACKBOARD_API_KEY, docs at .
Every step of the workflow, numbered, including which Backboard features each step uses (web search? memory? a specific model?).
Input and output: how the user gives input (command-line argument? question at start?) and exactly what files come out.
Quality bars: what the output must include (sources, sections, counts).
The definition of done: build it, run it on a real example, show the output, print usage instructions.

Example one-shot prompt (Research Reporter)

Study how specific this is. Then write your own, do not just copy it. Judges (and your own pride) can tell.

Build a complete "Research Reporter" agentic workflow, then run it once to
prove it works. One command runs the whole pipeline.

Tech requirements:
- Python, using the backboard-sdk pip package (Backboard API docs:
  https://docs.backboard.io).
- API key from the BACKBOARD_API_KEY environment variable, never hardcoded.
- Create ONE assistant named "Reporter" (save its id to reporter.json and
  reuse it on later runs).

The workflow, run by: python report.py "any topic here"

Step 1 (GATHER): Send the topic to Backboard with web_search="Auto" and ask
for the 5 most important recent facts and developments about it, each with
its source.
Step 2 (ANGLE): Feed those facts back to the assistant and have it identify
the 3 most interesting angles or takeaways, explaining why each matters.
Step 3 (WRITE): Have the assistant write a 400-600 word report in markdown:
a catchy title, an intro, one section per angle, and a sources section
listing the links found in Step 1.
Step 4 (SAVE): Save the finished report as report-<topic-slug>.md and print
the file path plus a 2-sentence summary to the terminal.

Rules:
- All steps run in ONE thread so the assistant keeps context between steps.
- Print a progress line as each step starts ("[1/4] Gathering...").
- If any API call fails, retry it once, then exit with a clear error message.

Definition of done: run the full pipeline on the topic "AI agents" and show
me the finished report-ai-agents.md contents, then print how I run it with
my own topic.

Fire your shot

New folder:

mkdir workflow-challenge && cd workflow-challenge

Make sure your key is set:
```
export BACKBOARD_API_KEY="your_key"
```
Start backboard (or open the folder in Studio).
Paste your prompt. Press Enter. Hands off the wheel.
Watch it build, run, and demo your workflow.

If the result misses the mark: figure out which sentence of your prompt let it happen. Vague step? No quality bar? Missing edge case? Fix the prompt, start a fresh session, and fire again. Re-rolls are free. Each one makes you better at the skill this whole challenge is about.

How to complete this challenge

Submit:

Your one-shot prompt (paste the text).
A screenshot or recording of the workflow running its steps and the finished output file.

Bonus points for creative workflow ideas we did not list.

Troubleshooting

Problem	Fix
Output quality is meh	Add explicit quality bars to the prompt: word counts, required sections, “each fact must have a source.”
Workflow forgets earlier steps	Require all steps to share one thread (context lives in the thread).
Auth or credit errors	Re-export `BACKBOARD_API_KEY` in this terminal, check balance at app.backboard.io.
Build agent stopped halfway	Your prompt likely lacked a “definition of done.” Say exactly what proves it works, and require it to run the demo.

Next up → Challenge 6: One-Shot a Multi-Agent Web App. The final boss. Same rules, one prompt, but now it is a whole web app where multiple agents work together. Everything from Challenges 4 and 5 was training for this.