Quiet Competence Beats Loud Waste

A dispatch on context engineering. Your corporation handing out unlimited tokens does not mean you have to spend them.

On June 10th, Nick Hodges published The tokenmaxxing backlash is coming in InfoWorld: developers are burning tokens at industrial scale on agentic coding; output volume is up, output value unclear; governance is about to land on the practice from outside because practitioners didn’t produce discipline from inside fast enough. He’s right — and the discipline is already practicable. It doesn’t look like governance. It looks like personal proficiency built on a toolkit most coders don’t know exists yet.

The through-line (an inversion worth naming). In food systems, individual responsibility gets weaponized to deflect from structural waste. AI is the opposite shape: the corporate layer hands out unlimited tokens, and the person at the keyboard is the only operator positioned to make a real reduction decision. The lever that works is the one the corporate frame suppresses — personal proficiency. Same output, fewer tokens. In both, the practitioner is the only one with line of sight to the actual problem.

Three years in, we know more. The shift from prompt engineering to context engineering isn’t a rebrand. Prompting is asking the right question. Context engineering is loading the model with the substrate it needs before you ask — roughly an order-of-magnitude token difference. The 2022 toolkit (single-shot queries, perfect-prompt mythology, iterate-until-it-works, AI-as-oracle, token spend unchecked) gives way to the 2026 toolkit (substrate loaded before query, model knows your conventions, two-to-three cycles not twenty, AI as thinking partner, token spend visible and tracked). Volume is not value. Tokenmaxxing is the supermarket pattern applied to AI.

Three moves: (1) Notice what you’re actually spending — open the session token counter, note your baseline. (2) Load context before you burn cycles. (3) Make quiet competence visible where it counts. And the toolkit is wider than you think: open-source models, local inference, model-by-task routing, fine-tune-small-route-to-frontier-for-edges, sovereignty as architecture.

Why it matters: every AI iteration consumes data-center water and energy, often in the same bioregions where regenerative work operates. Personal proficiency reduces that pressure at the only point in the system where reduction is voluntary. The climate connection is downstream of competence, not upstream — you don’t have to care about watersheds to do this; you just have to want to be good at your work. The downstream effect lands either way. (CC BY 4.0 — fork it for your domain.)