Red Shell · Common Errors
1 / 14

Project updates: what’s next

Overview
PDF930 KB PPTX16.5 MB
Slide 1 of 14: Project updates: what’s next
Real examples on this slide

← → step through · Home End first and last · F full screen · each slide has its own link, #1 to #14

READ THIS FIRST · PROJECT UPDATES, 24 SEP 2026 Project updates: what’s next The latest client feedback sets three core requirements that cannot be missed from now on. Where anything later in this course says otherwise, these updates apply. 1 Enforce the 80/20 rubric split At least 80% of positive points · Task Completion Did the agent produce the required artifact, follow the instructions and deliver the outcome the user asked for? 20% at most Anything else: tool use, process, agent narration Whenever possible, avoid process and tool use rubrics altogether. The client specifically called this out as important. A SIMPLE RULE A criterion on how the agent got there belongs in the 20%. A criterion on what the agent actually delivered belongs in the 80%. When in doubt, write the Task Completion criterion instead. TWO MORE FIXES FROM THE SAME REVIEW × Never write a rubric that only tests whether a file exists. × Every literal a criterion references matches the actual attachment exactly. One rubric cited insta.tiff when the file was klin.png. 2 Follow the new use case taxonomy The client has shared additional categories and subcategories, and tasks have to follow them much more closely. Keep every task clearly aligned with its intended use case, without drifting into other categories. Specific tool expectations may follow for some scenarios, to keep tasks aligned with the taxonomy. 3 The next 200 tasks are single turn The next block is 200 single turn tasks, rather than multi turn. Several open items are still being addressed on multi turn, so the client wants to move some volume through single turn tasks. CLIENT FEEDBACK · 24 SEP 2026 READ THIS FIRST