01/09/2026
When prompting coding agents, conversational politeness matters far less than specificity. The workshop encouraged participants to focus on precise instructions instead of filler language.
Here is what we learned from the AI Product Sprint at Imaguru Startup HUB, led by Andrew Tomin (Senior Product Manager, Agentic Ops) and Vitalijs Visnevskis (AI Product & Technology Expert): directing LLMs or coding agents requires carefully guiding the model through a structured workflow.
Here are the core ground rules for running effective product sprints with AI:
Rule 1: Plan First, Build Second
Do not ask an LLM to generate code or a full application right away. First, force the model through evidence-based research, generate a lightweight PRD (Product Requirements Document), and define the stack. Only after creating a clear specification should you instruct an agent or executor to build. The PRD exists to instruct the coding agent.
Rule 2: Be Specific, Not Polite
Remove unnecessary conversational filler and state requirements directly. State precisely what you want, what platform it is for, and what technical constraints apply.
Rule 3: Separate Evidence from Assumptions
Clearly state whether a detail is a confirmed insight or just a suspect assumption. Label assumptions explicitly so later validation can confirm or reject them. LLMs tend to average information, so guiding the narrative with clear parameters prevents vague outputs.
Rule 4: Manage Context Windows & Model Speeds
When rapid iteration matters, avoid unnecessarily slow reasoning modes. For workflows that exceed a model's context window, create separate chats, especially for local models or long conversations, while carrying forward only the relevant outputs.
From PRD to Working Prototype
During the workshop, teams following the discovery → validation → PRD workflow moved from ideas toward working prototypes in under two hours. Participants built live MVPs, including an OCR-based receipt spend analyzer, a location-verification platform for video production, and a browser extension for restaurant reservations.
As Andrew and Vitalijs emphasized during the session: The final responsibility for the output, code, and context always lies with the human builder, not the model
