Custom Instructions that survive model updates

Most Custom Instructions are written once, against one model, and quietly rot. A model update changes default behavior and suddenly your carefully tuned setup produces something different. The fix is to write instructions that describe intent and constraints, not model quirks.

Why instructions break

People over-fit. They discover a model ignores a request unless you SHOUT IN CAPS, so they bake caps everywhere. The next model doesn't need the caps — now your instructions are noisy and brittle.

The durable structure

Write instructions in three layers that any model will respect:

1. Identity — who Claude is for you. "You are my chief of staff."
2. Defaults — how to respond unless told otherwise. "Be concise. Lead with the answer."
3. Hard constraints — non-negotiables. "Never fabricate citations. Flag uncertainty."

# Identity
You are my writing partner for a B2B SaaS blog.
 
# Defaults
- Lead with the takeaway, then support it.
- Plain language. No marketing fluff.
- Markdown, with short paragraphs.
 
# Constraints
- Never invent statistics or sources.
- If a claim needs a source I don't have, say so.

Test it the right way

After any model update, run the same three prompts you always use and compare. If output drifts, fix the instruction, not the prompt. Durable instructions need editing once a quarter, not once a week.