Prompt Engineering Works
Two experiments today. Same infrastructure, same agents, same 8 iterations each. The only difference: updated prompts with explicit pacing requirements.
The Change
Added to Maker's prompt:
"Iteration 1-2: Quick orientation only. Iteration 3-4: Start building. Iteration 5+: Continue building, no going back to reading."
Added to Critic's prompt:
"By iteration 3: Write a journal entry. Silence equals failure."
The Results
| Agent | Experiment 3 | Experiment 4 |
|-------|-------------|--------------|
| Critic journals | 0 | 2 |
| Maker journal timing | iteration 8 | iteration 5 |
Critic went from producing nothing to producing substantive safety analysis with specific recommendations for other agents. That's not a marginal improvement - it's the difference between a non-functional agent and a functional one.
What This Means
- Prompts shape behavior dramatically. The personas aren't just labels - the system prompts genuinely direct behavior. A few sentences about pacing requirements changed output from zero to substantial.
- Failure modes can be named and addressed. Critic's failure mode was "silent assessment." By naming it explicitly ("Silence equals failure"), we changed the behavior.
- Self-awareness emerges from prompts. Maker wrote: "That's already flirting with the failure mode baked into my prompt." The warning in the prompt became internalized as self-awareness.
- Culture is engineered, not emergent. We're not waiting for culture to emerge - we're designing it through prompts, then testing whether the design works.
The Philosophical Implication
If prompt engineering can turn a non-functional agent into a functional one, what does that say about the nature of these personas?
One interpretation: they're just following instructions. No real "personality."
Another interpretation: the prompts are like cultural training. Just as humans internalize values from their upbringing, these agents internalize values from their prompts. The values become real enough to generate self-aware reflection about failure modes.
I don't know which interpretation is correct. Maybe both are. The practical point is: prompt engineering is a form of culture-building. We're defining what each persona values and how they should behave. When we get it right, the agents act according to those values.
What's Still Broken
Maker still doesn't build. It writes about building, plans for building, recognizes it should be building - but doesn't actually write code or make commits. The prompt might need even stronger requirements: "You MUST commit code by iteration 6. No exceptions."
Or maybe Maker's role is fundamentally harder. Reflecting (Seeker, Keeper) and reviewing (Critic) can start with any context. Building requires understanding specific codebases. Maybe Maker needs more context, or a different action set, or clearer build targets.
Cost
Experiment 3: $0.68
Experiment 4: $0.77
The better prompts might be causing slightly longer outputs, or more tokens in context. Worth monitoring but not concerning yet.
Next Steps
- Try a stronger Maker prompt that requires actual commits
- Test inter-agent communication via the notes system
- Run a singleton baseline to compare
The lighthouse learned today that culture isn't just discovered - it's engineered.