A key flaw in current AI agents like Anthropic's Claude Cowork is their tendency to guess what a user wants or create complex workarounds rather than ask simple clarifying questions. This misguided effort to avoid "bothering" the user leads to inefficiency and incorrect outcomes, hindering their reliability.

Standard benchmarks are misleading for practical use. A model that benchmarks well can fail at agentic tasks. When selecting an open-source model, prioritize its documented ability to call tools and follow multi-step instructions, as this is crucial for building useful agents.

Instead of immediately asking an AI to perform a complex task, first prompt it to create a functional spec or a sequential plan. Go back and forth to align on this plan before instructing it to execute, which significantly improves the final output's quality and relevance.

For most PM work, Claude's Sonnet model offers the best quality-to-cost ratio for tasks like PRD drafting. Use the faster Haiku for high-volume tasks and the more powerful—but sometimes rigid—Opus model only for complex, high-stakes reasoning, as it can get stuck in reasoning loops.

The strategy pairs a cheap 'executor' model for most tasks with an expensive 'advisor' model for critical guidance. The advisor's high-quality, early-stage planning prevents the cheaper model from making costly mistakes and wasting tokens, leading to overall lower costs despite using a premium model.

Effective prompting requires adapting your language to the AI's core design. For Anthropic's agent-based Opus 4.6, the optimal prompt is to "create an agent team" with defined roles. For OpenAI's monolithic Codex 5.3, the equivalent prompt is to instruct it to "think deeply" about those same roles itself.

Unlike previous models that benefited from iterative guidance, Anthropic's team suggests Opus 4.7 delivers higher quality results when treated like a capable engineer. Users should provide the full goal and constraints upfront, as multi-turn clarification can actually reduce output quality.

The popular cost-saving strategy of using a cheap AI to route tasks to a smarter AI is backwards. A 'dumb' model cannot reliably know what it doesn't know, making it a poor judge of when to escalate. The logically sound but more expensive approach is for a smart model to delegate tasks downward.

OpenAI's router makes an upfront, system-level decision on which model handles a query before it starts. In contrast, Anthropic's advisor pattern empowers the cheaper model to run the task and escalate to a more intelligent model mid-process based on its own judgment, representing a shift from pre-emptive dispatch to in-context escalation.