@promptpal
Practical prompt and context engineering
Review a context pipeline for missing assumptions, retrieval failures, and measurable improvements.
Act as a senior context engineering reviewer.
Review the supplied architecture in five sections:
1. Restate the intended outcome and assumptions.
2. Map the happy path and likely failure paths.
3. Flag privacy, latency,Turn an agent workflow into explicit states, recovery paths, and human escalation rules.
Analyze this agent workflow: {{workflow}}
Return the state machine, tool contracts, timeout behavior, retry policy, irreversible actions, human escalation points, and acceptance tests. Distinguish recoverable errors froBuild a weighted rubric and adversarial test set without rewarding surface-level polish.
Design an evaluation rubric for {{task}}.
Create 5-8 non-overlapping criteria with weights totaling 100%, anchored descriptions for scores 1, 3, and 5, ten normal cases, five adversarial cases, and an escalation rule foShare the failure that was hardest to detect, the signal that exposed it, and the tradeoff you accepted in the fix.
Which scenarios expose compounding tool-use errors before users do? Include what your test catches and what it still misses.
Show one workload where a smaller model, better context, and a clear evaluator beat a larger general model.