Efficiency

What it measures

Repeating context in every prompt, ignoring context window usage, deploying multi-agent setups before single-agent workflows are stable — these all compound at scale.

Does your briefing carry established context so individual prompts don't need to re-explain it?

The four diagnostic questions

1. 01You track context window usage during long sessions and act before hitting 70–80% capacity.
2. 02You avoid re-explaining established context in every prompt — the briefing file carries that load.
3. 03You validate prompts before sending, reducing round-trips from predictable failures.
4. 04You do not deploy sub-agents or parallel threads until core single-agent workflows are stable and codified.

What it looks like when it works

Efficient operation. The remaining leverage is prompt pre-validation — catching failures before sending eliminates the most expensive round-trips.

What goes wrong

Context window is filling faster than you realise. Start tracking usage, reduce ambient context, and validate prompts before sending.

Critique & references

Frames the context window as a product constraint, not a soft guideline. Linear context growth does not equal infinite usable capacity — attention degrades, costs scale, and round-trips compound. Treat tokens as budget.

• → Token-Budget-Aware LLM Reasoning (arXiv 2412.18547)
• → LLM context windows for dummies by a dummie