AI Budget: More Jazz, Less Math
1. Context
I tried to frame AI context memory like a staffing budget: neat percentages, clean splits. My PM brain loved it. But in practice, it behaves less like math and more like music. The trick isn’t perfect allocation, it’s finding the right rhythm between backbone, values, and expertise. And while this metaphor sounds playful, the discipline behind it is real: designing agents/companions isn’t about the character count math, it’s about making intentional choices that shape outcomes.
2. What I Tried
I broke the AI context window into three roles:
- BIOS = Backbone (always on)
- CultureCore = Value System (helps with judgment calls, sometimes even overrides BIOS)
- Domain Expertise Packs = Knowledge Add-Ons (specific fields you plug in as needed) At first, I even gave it percentages (40/25/35) as if it were a strict budget.
3. What Happened
It sounded smart in theory, but the numbers made it feel like math when it’s really more art. My default PM brain wanted tidy percentages, but agents/companions don’t work that way. As I prototyped, I realized adding or removing KnowledgePacks shifted behavior far more than any “40/25/35” split could explain. The model responded more fluidly than a hard budget ever would.
4. Takeaway
Keep the staffing analogy, but simplify:
- BIOS = backbone (always there)
- CultureCore = value system (guides and sometimes overrides)
- Domain Expertise = specific knowledge (only when it adds value) It’s not about dividing up a finite pie. It’s about design discipline and judgment.
5. Adapt This
When you’re shaping an AI companion:
- Always include BIOS + CultureCore. That’s your baseline.
- Add Domain Expertise Packs selectively. But treat it as an art, not a formula. My mistake was overloading a companion with process-heavy packs (vetting steps, operations flow). Instead of useful strategy, it spat out generic reminders. Once I removed that pack, the companion naturally connected campaigns and brand strategy again. Voila, no instruction changes needed. Rule of thumb: the more you cram in, the less room the companion has to think with you.
6. Reflection
AI context memory isn’t about precise allocation, it’s about integration: bringing backbone, values, and expertise together in the right proportions so the team has the most relevant companion for the project. The real skill is choosing resources so the system amplifies strategic work rather than weighing it down. I can’t predict when or if this skill will be replaced by AI. What I do know is that I’ll keep experimenting to see how these tools can support teams: showing up with less noise to fight through and more space for the fun challenges. That curiosity, even with a little anxiety, is what keeps me going.