Anyone can paste an instruction block. Few design one.
12 design patterns pulled from building 52 real Copilot Chat agents, in production and open source. Each pattern is a before/after block and the exact failure mode it prevents.
“Battle-tested in production environments.”
Jared Spataro, Microsoft’s CMO for AI at Work, citing the author’s public Copilot prompt library on LinkedIn. This guide applies the same standard: everything in it has been run for real, and the failures are documented next to the wins.
Most blocks aren't wrong about the job. They're silent about the rest.
An instruction block rarely fails because it misdescribes what the agent should do. It fails on everything around that, the parts a quick draft leaves out.
Silent on what it must never do. The block says what to do, and says nothing about the action that would make it harmful.
Vague on edge cases. It assumes clean input. Real users paste half a thought, and the agent guesses.
No quality check. Nothing catches a bad output before it ships to the 50 people who @mention the agent.
"Be helpful." It states a vibe instead of naming the exact action that would count as unhelpful.
The craft behind 52 agents, written down.
Not "prompt tips." Named patterns, named failure modes, a budget, a scaffold, and full rewrites you can copy.
12 Design Patternsbefore / after
Each pattern is a before/after instruction block plus a plain explanation of the failure mode it prevents.
Failure Mode Catalogue8 named
Eight named failures, each with a detection test so you can spot it, and a concrete fix.
The 8,000-Char Constraintbudget
The hard ceiling on a Copilot Chat instruction block, broken into a section-by-section character budget.
The Agent Scaffoldreusable
A copy-pasteable template with the sections in order, so a new agent starts from a structure, not a blank box.
5 Full Worked Examplescomplete
Complete instruction-block rewrites across five different agent types, start to finish.
10-Question Test Checklist+ budget audit
A deployment testing checklist, plus a worked character-budget trim from 9,200 down to under 8,000.
8,000 characters. Every section competes for them.
A Copilot Chat agent instruction block has a hard character ceiling. Spend it badly and you cut the quality checks to fit the role description. The guide gives you the budget, and a worked trim from 9,200 characters to under the line.
The guide's worked audit trims a real 9,200-character block to under 8,000 without dropping a single guardrail.
Anyone whose agent gets @mentioned by more than one person.
If you are building Copilot Chat agents that need to behave consistently at scale, not just for you but for everyone on your team who calls them, this is the design layer underneath that consistency.
Stop pasting. Start designing.
- 12 patterns, each a before/after block
- 8 named failure modes with tests and fixes
- The scaffold, the 8K budget, and 5 full rewrites
- A 10-question deployment test checklist