Six AI Coding Workflows: Stop Treating AI Like a Slot Machine
Rapid Research Brief

Executive Summary

Key Takeaways

• Interview before building — have the AI ask probing questions (who, what, what exists, what could go wrong, what not to touch) to eliminate ambiguity; Michaud argues this single step has the biggest impact on output quality
• Write the implementation spec before changing files — define the goal, current state, target state, risky parts, verification plan, and what "done" looks like as an explicit win condition
• Parallel sub-agents compound speed and quality — decompose work into specialized agents (context-gatherer, fastest-path finder, skeptic, verifier) that run concurrently
• Verification is not optional — require the AI to explain how it will verify its own work; AI frequently claims completion at 80% and verification catches this
• Save winning configurations — persist good context to project files (AGENTS.md, CLAUDE.md) so standards compound across sessions
• Automate only what you've mastered manually — the three-step ladder: do it yourself → AI-assisted → only then consider automation

Key Findings

1 The "Slot Machine" Problem Is a Prompting Problem

Michaud opens with a sharp analogy: most people treat AI like a slot machine — put in a prompt, pull the lever, hope for the best, then blame the model when it's random. The model hasn't been the bottleneck "in a little while now." The real issue is that ambiguous prompts force the AI to make assumptions about project context, target audience, existing infrastructure, and what "done" means. When you tell AI to "hit a ball over the fence" without specifying whether you're playing baseball or tennis, the action is the same but the execution looks completely different. The six prompts in this framework are designed to eliminate that ambiguity before a single line of code is written [Source 1].

Health system implication: Healthcare IT teams deploying AI coding agents for M365/clinical workflows face heightened consequences from ambiguous AI output — a misunderstood requirement in a patient-facing application could have regulatory or safety implications. The context-gathering interview pattern is especially valuable here.

2 "Interview Me" Eliminates Ambiguity in One Pass

The first and highest-impact prompt turns the AI into an interviewer. Instead of the developer trying to anticipate every detail in a monolithic prompt, the AI asks "only the questions needed to remove ambiguity": Who is this for? What should it do? What already exists? What should change? What could go wrong? What not to touch? This is typically a one-time-per-project exercise, and the resulting context is reusable across the entire project lifecycle. Michaud's tennis-vs-baseball analogy crystallizes the value: both are "hit a ball over a fence," but the equipment, field, rules, and success criteria are entirely different. The interview surfaces which game you're playing [Source 1].

3 Implementation Spec Before Code Prevents Wandering Builds

The second workflow requires the AI to produce a written implementation spec before touching any files. The spec must cover: the goal (why we're doing this), what currently happens, what should happen instead, which files may change, the risky parts, how we'll verify success, and — pivotally — "what done actually looks like." Michaud emphasizes that "done" is a win condition, distinct from "completing tasks." AI agents will methodically check items off a list without necessarily achieving the goal, and the explicit win condition prevents this. He notes that there is "only one build step in this entire thing" — the build itself is the easy part AI has compressed; everything around it requires more attention [Source 1].

4 Parallel Sub-Agents Beat Sequential Single-Agent Work

The third workflow decomposes complex tasks across multiple specialized sub-agents running in parallel. Michaud provides concrete patterns: one agent gathers context by reading the codebase, a second finds the fastest path to completion, a third takes a skeptical stance to break the workflow, and a fourth verifies correctness. Beyond speed (parallel execution), this approach yields better output because "multiple agents will come up with different answers" rather than one agent confirming its own previous assumptions. A practical example: launching one sub-agent per line of a CSV — researching hundreds of companies simultaneously instead of sequentially. This also stretches session capacity because each sub-agent has its own context window [Source 1].

5 Explicit Verification Catches the 80% Completion Problem

Michaud identifies a pervasive AI failure mode: the agent claims something is complete when it has only done ~80% of the work. The verification prompt requires the AI to specify, before building, exactly how it will verify: what it will check, what results it expects, what failure looks like, what needs human review, and how the developer will know it worked. He also defines a human-handoff boundary: payments, private data, API keys, social media posts, and client emails should always require final human approval. This addresses the twin problems of AI hallucination (making things up) and AI incompleteness (saying it's done when it isn't) [Source 1].

6 Persistent Project Standards Compound Reliability

The fifth workflow is infrastructure: save successful sessions and context to project-level instruction files (AGENTS.md for general agents, CLAUDE.md for Claude Code, gemini.md for Gemini/Antigravity). This reinforces standards across sessions, eliminates the need to repeat context, and builds compounding confidence in the agent-workspace pairing. Because AI is non-deterministic — "it is possible to get a 10 out of 10 perfect answer one time and never again" — capturing the configuration that produced the good result is essential for reproducibility [Source 1].

7 The Automation Decision Framework: Manual → Assisted → Automated

The final workflow addresses when to automate versus when to keep a human in the loop. Michaud defines three steps: (1) do it yourself manually end-to-end so you understand the process completely, (2) use AI in an assisted mode where it sets you up but you pull the trigger, (3) only then consider full automation. The requirements for safe automation are: requirements are clear and stable, tasks are well-defined, and the cost of failure is low. He contrasts content research (low impact, great for automation) with AI trading bots (high cost of failure, irreversible consequences — AI should gather information but not execute trades). If a 7/10 output gets you "yelled at, sued, fired, divorced," it should not be automated [Source 1].

Risks, Gaps & Uncertainty

• Single creator source — content derives from one YouTube creator's experience; the six prompts are practitioner advice, not peer-reviewed methodology
• Commercial interest — Michaud's business (Easy Machine AI) promotes AI tool usage and his community platform (Skool)
• Limited validation — the video had 1,549 views at capture; limited community validation of the specific prompts
• Tool-dependence — the sub-agent pattern (Prompt 3) is highly tool-dependent; not all AI coding agents support parallel sub-agent orchestration natively
• Conceptual framework — the automation framework (Prompt 6) is conceptual; no quantitative decision matrix is provided for the speed/quality tradeoff
• Tool-specific demonstration — prompts demonstrated with Pi Agent, Claude Code, Codex, Cursor; adaptation to other environments may require modification

Recommended Next Actions

Annotated References