Part of #11
Question
How are agent instructions authored, versioned, and tested?
app/instructions/coach_agent.md is 163 lines and most of it is negative: do not call an item a T4 finisher unless verified, do not say "if you mean your own games", do not repeat the same fact in the opening sentence and the bullet list and the closing sentence. Each line is a real failure someone patched by hand. The file has become a changelog of bugs, and nothing tests whether any given line still earns its place.
Decide:
- what belongs in an instruction versus in code, a typed contract, or a tool's own description
- how a behavioural rule gets added — what evidence justifies one, and what retires one
- how instructions are structured so a rule can be found, changed, and traced to the failure that motivated it
- whether formatting rules belong in the prompt at all, or in post-processing
- how a change to instructions is validated before it ships
Depends on the grounding contract and the topology: both determine what instructions still have to carry.
Part of #11
Question
How are agent instructions authored, versioned, and tested?
app/instructions/coach_agent.mdis 163 lines and most of it is negative: do not call an item a T4 finisher unless verified, do not say "if you mean your own games", do not repeat the same fact in the opening sentence and the bullet list and the closing sentence. Each line is a real failure someone patched by hand. The file has become a changelog of bugs, and nothing tests whether any given line still earns its place.Decide:
Depends on the grounding contract and the topology: both determine what instructions still have to carry.