Workflow · Production

Building Your Own Product Copilot: Challenges, Opportunities, and Needs

The problem

Software engineers building product copilots face pain points at every stage: prompt engineering is fragile and time-consuming, orchestrating multi-step AI workflows is hard to steer, and existing software engineering processes have not caught up with the demands of building AI-powered applications.

First attempt

Models produce inconsistent, hallucinated, or malformed outputs; get stuck in loops; and mistakenly signal task completion prematurely, while unit-testing frameworks designed for deterministic code cannot cope with generative model variability.

Workflow diagram · grounded in source
1
User natural-language query
trigger
“the initial scope of work involved supporting single-turn interactions, where the user would provide a query or command, and the copilot would provide a response”
2
Intent detection
ai_action
“we first send that query to the copilot and ask like what kind of intent does the user have for this specific query out of intents that we redefine and provide”
3
Skill routing
routing
“Once an intent is detected, the prompt is then routed to the appropriate skill, "like adding a test or generating documentation," that is capable of handling the request”
4
Dynamic prompt assembly
ai_action
“these components "gets populated and modified before the final query"”
5
Output parsing
output
“After the model returned a response for the prompt, additional processing was necessary in order to interpret the response”
6
Human review of AI output
human_review
“this content is AI generated, and you know you should review all of it before you decide to do anything further”
Reported outcome

The study produced collaborative findings and tool design directions for the software engineering community, with participants converging on prompt asset management, orchestration tooling, and AI-specific benchmarking as the highest-priority unmet needs.

Reported metrics
Prompt engineering effortextremely time-consuming and resource-constrained
Test runs per test case10
Passing threshold per test case7 of the 10
Labeled responses for benchmark dataset10k
Show all 5 reported metrics
prompt engineering effortextremely time-consuming and resource-constrained
test runs per test case10
passing threshold per test case7 of the 10
labeled responses for benchmark dataset10k
per-test LLM inference cost1-2 cents
Reported stack
GPT-4ClaudeLLaMAGitHub CopilotMural
Source
https://arxiv.org/html/2312.14231
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The study produced collaborative findings and tool design directions for the software engineering community, with participants converging on prompt asset management, orchestration tooling, and AI-specific benchmarking…

What tools did this team use?

GPT-4, Claude, LLaMA, GitHub Copilot, Mural.

What results were reported?

Prompt engineering effort: extremely time-consuming and resource-constrained; Test runs per test case: 10; Passing threshold per test case: 7 of the 10; Labeled responses for benchmark dataset: 10k (source-reported, not independently verified).

What failed first in this deployment?

Models produce inconsistent, hallucinated, or malformed outputs; get stuck in loops; and mistakenly signal task completion prematurely, while unit-testing frameworks designed for deterministic code cannot cope with ge…

How is this workflow AI workflow structured?

User natural-language query → Intent detection → Skill routing → Dynamic prompt assembly → Output parsing → Human review of AI output.