Instruction Budget: The Ceiling Moved

~200

Old rule of thumb (2025): instructions start dropping

~2,000+

Frontier (2026): named constraints tracked in one prompt

The question shifted from “can it follow this?” to “is it worth the tokens, and how do we prove it complied?”

Stop over-optimizing for “short prompt.” Optimize for structure + checklists.
Invest your time in verification: compliance checks + audit logs.
Length is now a cost/latency tradeoff, not a hard capability cliff.

Quick token budgeting proxy

# Count tokens as a consistent budget proxy (not exact for every vendor) python3 -m pip install --user tiktoken python3 - <<'PY' import sys, tiktoken enc = tiktoken.get_encoding('o200k_base') print(len(enc.encode(sys.stdin.read()))) PY < SKILL.md

Use this to compare skills/policies relative to each other. The goal is consistent budgeting, not perfect vendor token parity.