Eric Michaud presents six structured prompts that transform AI coding agents from unreliable "slot machines" into consistent development partners. Only one step involves actually building — the rest is the surrounding discipline: context, planning, decomposition, verification, persistence, and knowing when to automate.
AI asks probing questions before building. One-time per project, reusable context. Biggest single impact.
Written plan before changing files. Goal, risky parts, verification, what "done" looks like.
Context-gatherer, fast-path finder, skeptic, verifier. Faster + diverse perspectives.
Catch the 80% problem. Save to AGENTS.md. Automate only what you've mastered manually.
Key insight: Only one of the six steps is actually building. Everything else is the professional discipline that surrounds the code.
The AI becomes the interviewer, asking only the questions needed to remove ambiguity:
This is a one-time-per-project exercise. The resulting context is reusable across the entire project lifecycle. Michaud says this single step has the biggest impact on output quality.
The AI must produce a written plan covering:
"What done looks like" is a win condition, not just task completion. AI agents check items off a list without achieving the goal. The explicit win condition prevents this.
Reads the codebase and reports back so you don't waste the main context window.
Finds the quickest route to completion — minimal implementation.
Tries to break the workflow. Finds edge cases and failure modes.
Confirms correctness independently. Multiple agents = different answers, not self-confirmation.
CSV pattern: Launch one sub-agent per row — research hundreds of companies in parallel instead of sequentially.
AI frequently says it completed something when it only did ~80% of the work. The verification prompt catches this gap.
Before building, the AI must specify:
Human handoff boundary: High-impact actions always require final human approval.
AGENTS.md for general agents, CLAUDE.md for Claude Code, gemini.md for GeminiAI is non-deterministic: "It is possible to get a 10/10 perfect answer one time and never again." Capturing the configuration that produced the good result is essential for reproducibility.
This builds compounding confidence in the agent-workspace pairing over time.
The test: If a 7/10 output gets you yelled at, sued, fired, or divorced — do not automate it.
Good for automation: content research (low impact). Bad: AI trading bots (irreversible consequences).
Adopt "Interview Me" as project init step. Add to standards repo as pre-build checklist for new AI coding sessions.
Create AGENTS.md template. Fill-in-the-blanks: goal, current state, target state, risky parts, verification, done condition.
Evaluate prompts as Hermes skills. "Interview Me" and "Impl Spec" are strong candidates for slash commands or cron pre-checks.
Audit workflows against automation ladder. Identify processes that jumped straight to automation without manual mastery.
Add verification prompt to pre-deploy validation. Michaud's "what failure looks like" enriches existing quality gates.
Tier guide: Tier 1 = Primary institutional source (official website or announcement) · Tier 2 = Company-owned media or trade publication · Tier 3 = Breaking news / single-source — verify independently
Sources: YouTube video content. Full transcript captured via Obsidian Web Clipper (Eric Michaud / Easy Machine AI). Verify key claims independently.
Generated June 22, 2026 · DeepSeek V4 Pro · research-rapid v1.0 · Gambit / brief-to-slides v2.0