How I'm Actually Using AI Right Now (Including Claude Duet)

Context

This is a solo episode from NLW about how he is actually using AI right now across strategy, research, copywriting, image generation, data analysis, coding, and personal agents. The useful lesson for PMs is not the exact stack. It is the operating logic behind it: match models to specific jobs, build repeatable pipelines instead of isolated prompts, and judge personal agents by whether they save real work after setup and maintenance. The comparison between ChatGPT Dots and Claude Duet is especially relevant because it shows where proactive agents are starting to feel useful, and where they still create more management than value.

The Big Idea

AI becomes materially useful when you stop looking for one best model and instead design a stack around distinct jobs, risk levels, and workflow bottlenecks. Personal agents only matter if they either hold context better than normal chats or do proactive work with less upkeep than they create.

NLW's own workflow makes this concrete. Strategy lives in one model, research gets cross-checked across two models, building happens in coding tools, and agent products are judged by whether they remove work instead of becoming one more system to manage.

Key Insights

Match models to jobs

• NLW breaks his AI use into distinct work types: strategy, research, copywriting, image generation, data analysis, building, and proactive assistance.
• His core point is that model choice is now a fit problem, not a single leaderboard problem.
• For PMs, this is a better adoption pattern than choosing one company standard and pushing every task through it. Discovery, synthesis, stakeholder writing, and execution need different strengths.

Strategy needs real pushback

• NLW says roughly 60% of his AI time goes to strategic externalization, which means thinking out loud, pressure-testing ideas, and deciding what to do next.
• The failure mode for this job is sycophancy. A model that treats every idea as brilliant is barely better than writing on paper.
• He prefers Opus 5.5 here because it gives critical engagement. He found recent GPT models too flat for this kind of work.
• For PMs, the lesson is that strategy support is not mainly about polished text. It is about whether the model can challenge assumptions without becoming random or needlessly combative.

Research needs verification loops

• Because NLW turns research into public claims, he optimizes for sourcing quality and low hallucination risk, not just speed.
• He often has Astra and Opus research the same topic, then review each other's work before he turns it into content.
• That extra pass matters most when the output will influence an external narrative, a leadership memo, or any artifact where source quality matters.
• For PMs, this is a practical upgrade from asking one model once. High-stakes product research should have a verification step, even if the second pass is another model rather than another person.

Pipelines beat isolated prompts

• NLW's biggest gains came from building end-to-end systems around recurring outputs.
• One example is the AI Daily Brief production flow. A transcript becomes shareable web assets, then social posts, and then a clipped 60 to 90 second video segment.
• He built similar support for a new long-form show, where research, transcript handling, companion essays, visuals, and clips all connect into one workflow.
• For PMs, the important idea is that AI value compounds when one raw input, such as a transcript or research dossier, feeds multiple downstream outputs.

Agents must earn upkeep

• NLW previously experimented with always-on agents in Open Claw, including a research agent and a chief of staff agent.
• He stopped using them because the juice wasn't worth the squeeze. The maintenance and correction cost outweighed the benefit.
• This is a strong reminder that automation cost is not just setup. It also includes monitoring, fixing mistakes, and reworking the system when the workflow changes.

Duet outperformed Dots

• Dots felt mostly useful as a context holder for important ChatGPT work. NLW did not find it very proactive, and he found the personality too flat for strong strategic back-and-forth.
• His clearest example was sharing about six hours of strategic conversation from Claude into his Dot, only to get a response that mainly acknowledged the information had been saved.
• Duet looked more promising because it immediately connected to Gmail, Slack, and Drive, proposed proactive projects, and drafted responses to emails he otherwise would have starred for later.
• In one case he sent Duet's draft directly. In another, the draft got him about 90% of the way there.
• The PM lesson is to separate two jobs that personal agents often blur together: holding context and taking initiative. A product can do one well and still fail at the other.

Mental Models & Frameworks

Job-to-model map

Use a simple map of work type to model strength instead of picking one default tool:

• Strategy: Opus 5.5, because NLW values pushback and critical engagement.
• Research: Astra and Opus together, because the work needs sources and cross-checking.
• Copywriting: Astra, partly because of long accumulated context on what he likes and because he finds the output concise.
• Data analysis: Astra inside Codex, because he likes the results and the style of reasoning there.
• Building: Claude Code with Opus 5.5, usually with high or extra effort settings.
• Personal agents: Dots for context holding, Duet for more work-focused proactivity.

This model is useful when a team is tempted to standardize too early on one vendor or one interface.

Cross-review research loop

• Run two models on the same question.
• Have each model inspect the other's sources, omissions, and weak claims.
• Only then convert the output into something publishable or decision-driving.
• Use this when bad sourcing would create real cost, such as a strategy memo, product narrative, or external communication.

Completion energy test

A tool is not valuable just because it lowers the effort to start. It also has to lower the effort to finish the last 10 or 20 percent. NLW says Fable felt different because it reduced that completion energy enough for him to build production pipelines, not just rough prototypes. This is a useful PM test for any AI tool that looks exciting in demos but leaves teams with half-finished systems.

Monothread context model

A monothread is one ongoing conversation that keeps broad context across tasks instead of splitting work into many separate chats or projects. The benefit is lower cognitive overhead because the user can work in a more stream-of-consciousness way. NLW likes this model because it fits his way of thinking better than heavy hierarchy. Use it when work is fluid and cross-cutting, but keep channels or projects when boundaries still help.

Proactivity value test

A proactive agent has to pass three checks:

• Real task: the agent notices work you actually would have handled later.
• Meaningful draft: the agent completes enough of the task that you save time.
• Low correction cost: the agent is easy to redirect when it misunderstands business context.

NLW's earlier Open Claw experiments failed this test. Duet's email drafting sometimes passed it.

Decision Principles

Principle: Optimize for fit

When: choosing a model or agent for a recurring PM workflow. Why: NLW's stack shows that the best choice for strategy, research, copywriting, and coding can differ materially, so a single default AI setup will often underperform.

Principle: Verify public-facing claims

When: using AI research in a leadership memo, product narrative, or any work that will be shared publicly. Why: NLW cross-checks Astra and Opus because source quality matters more than speed when an output will shape decisions or public understanding.

Principle: Kill weak automations early

When: a proactive workflow needs constant correction or supervision. Why: NLW abandoned earlier chief of staff and research agents because they created more upkeep than value. A mediocre automation can quietly become a tax on the team.

Trade-offs & Nuance

Monothread versus folders

• Monothread reduces cognitive load because you do not have to manually manage as many chats and projects.
• Separate channels or project spaces still help when you want hierarchy, cleaner boundaries, or more focused work.
• Duet is notable because it supports both. It has a main everything stream, but also separate channels when structure helps.

Proactivity versus false alarms

• The same initiative that makes an agent useful can also make it noisy.
• Duet flagged a sponsor inventory issue multiple times because it saw two sponsors offered the same date, but it did not understand that multiple ad slots existed on those days.
• PMs should not judge proactive agents only by whether they surface tasks. They should judge them by precision and how easily they can be corrected.

Unified agents versus best tools

• NLW does not want to do everything inside Duet.
• He found a UI and brand visual design task worked worse there than in Claude's dedicated design experience.
• The practical implication is that orchestration and creation may stay separate. A personal agent can coordinate work without being the best place to do the work.

Consumer help versus work help

• NLW remains skeptical about consumer agent use cases such as flight booking.
• His reasoning is straightforward. If preferences are simple and already constrained to one airline or one booking flow, explaining every rule to an agent can take as long as just doing the task directly.
• He is much more optimistic about professional workflows where context is scattered across email, Slack, Drive, and ongoing tasks.

Common Mistakes

Mistake: Chasing one best model

• NLW explicitly argues that the important question is usually not whether a new model fully replaces another one.
• The better question is where that model fits in the stack.
• Teams that standardize too early can miss better combinations, such as using one model for critique-heavy strategy and another for concise copywriting or data work.

Mistake: Automating vague workflows

• NLW's earlier always-on agents did not last because the workflows were not clearly valuable enough to justify maintenance.
• Proactive systems work better when the job is narrow and repeatable, such as drafting a response to an email he already would have handled later.
• If the workflow itself is fuzzy, the automation usually amplifies the fuzziness.

Mistake: Mistaking memory for initiative

• Dots was somewhat useful as a place to dump context, but that did not automatically make it a strong proactive assistant.
• A PM evaluating agent products should test whether the product only remembers things, or whether it takes the right next step without heavy prompting.

Practical Application

Map your AI stack

• List the top five recurring PM jobs in your week, such as roadmap strategy, user research synthesis, data analysis, stakeholder writing, and prototyping.
• For each job, write what good output actually requires: pushback, sourcing, concision, visual generation, or execution speed.
• Choose tools against those requirements, not by brand loyalty. NLW's stack works because each tool has a defined role.

Add a research cross-check

• For any product memo that will influence leadership or external communication, run a second pass that critiques the first model's claims, sources, and missing evidence.
• Treat the second pass as a reviewer, not as a second draft writer.
• This is a fast way to lower hallucination risk without turning every memo into a giant manual fact-checking exercise.

Build one reusable pipeline

• Pick one recurring artifact with downstream reuse, such as customer interview transcripts, sales call recordings, or release notes.
• Design a flow that turns that raw input into multiple outputs, such as a summary for the team, a problem list for discovery, and a short stakeholder update.
• NLW's biggest gains came from pipeline thinking, not from isolated prompts.

Pilot one proactive workflow

• Start with a narrow job that already happens regularly, such as drafting replies to starred customer emails or surfacing unresolved Slack decisions before a planning meeting.
• Measure three things for two weeks: time saved, number of false alarms, and how often you had to restate the rules.
• If the workflow fails on any of those, kill or redesign it quickly.

Keep specialized tools separate

• Use a personal agent for coordination only when it is clearly better at holding context or nudging action.
• Keep tasks like visual design, coding, or detailed data work in the dedicated environment that already performs best.
• This avoids the common trap of forcing an orchestration layer to become a creation layer.

Questions to Consider

• Which recurring PM task in your week actually needs critical pushback, and are you using a model that challenges assumptions instead of approving every idea?
• Where in your product workflow would a cross-review step, such as one model checking another model's sources and claims, materially reduce decision risk?
• Which proactive task in your email, Slack, or document workflow is narrow enough to automate, but painful enough that you would notice if it disappeared?
• Are you treating an AI agent's memory of project context as proof that it can also take the right next action, or have you tested those two capabilities separately?
• What product process would improve most if one raw artifact, such as a customer interview transcript, automatically fed three downstream outputs instead of one?

Bottom Line

The durable lesson is not to copy NLW's exact stack. It is to design AI around distinct jobs, add verification where mistakes are expensive, and demand that personal agents save more work than they create. The tools will converge quickly, but the PM advantage will come from clearer workflow design and stricter judgment about where automation actually earns its place.

Tools & Products