Module 5: Trust But Verify (The Critique Loop)

Pillar 04: "Never accept the first output."

What you need to know first

Plan vs. implementation: A plan is what you're going to build. Implementation is the actual code. Plans are cheap to change; code is expensive to change. SOLO insists you get the plan right before writing code.

Critique: Asking the AI to tear apart its own plan. This sounds strange — why would it find problems it just created? Because LLMs generate one plan by following the path of least resistance. When you explicitly ask them to attack that plan, they activate a different reasoning pattern and often find real issues.

Revise: Updating the plan based on critique findings.

The SOLO idea

The planning loop is: /architect → /critique → /revise → repeat if needed. You never implement the first plan. You always run at least one critique pass. The critique catches shortcuts, missing edge cases, and unstated assumptions.

But here's what the blog doesn't tell you: not every critique finding is worth acting on. Some are real problems. Some are the AI being pedantic. You need to learn to tell the difference.

Reading

• SOLO blog: "The planning loop"
• Open workflows/critique.md in VS Code and read it — note the list of "Anti-Patterns to Flag"
• Open workflows/revise.md and read it (it's short)

---

Lab 5: The Critique Gauntlet

Estimated AI conversations: 2-3

1. Get a plan. In Copilot Chat, type:

   Follow #file:workflows/architect.md — Here is my spec: #file:SPEC.md. Produce a detailed technical plan for building this.

   Save the AI's output as PLAN_v1.md.

2. Critique it. In the same conversation, type:

   Now follow #file:workflows/critique.md — Critique the plan you just produced.

   Save the output as CRITIQUE_v1.md.

3. Now, learn to evaluate the critique. Read every finding and label it:

   Label	Meaning	Example
   🔴 Real problem	Would cause a bug or break something	"No validation on date input — program will crash if user enters 'yesterday'"
   🟡 Good idea, not critical	Would improve the code but isn't a blocker	"Could add a confirmation prompt before deleting a session"
   ⚪ Noise	Pedantic, over-engineered, or irrelevant for this project	"Should implement a plugin architecture for future extensibility"

   Write your labels directly in CRITIQUE_v1.md next to each finding.

4. Revise the plan. In the same conversation, type:

   Follow #file:workflows/revise.md — Address these findings from the critique: [paste the 🔴 and 🟡 findings]. Ignore everything else.

   Save the output as PLAN_v2.md.

5. Run critique one more time on the revised plan. This second pass usually finds fewer issues. If it finds zero, you're done. If it finds real ones, revise again.

6. Commit everything:

   git add PLAN_v1.md CRITIQUE_v1.md PLAN_v2.md
   git commit -m "Planning loop: architect, critique, revise"
   git push
   

The habit to build: Never implement the first plan. Always critique. But also: learn to filter. The AI will generate 15 findings, of which 3 matter. Your job is to find the 3.