ExpertsFMClaude Duet is beautiful. I still forgot to use it
How I AI · October 21, 2026
5 Selected Lessons:
heuristic · 3 case · 1 reasoning_pattern · 1
Speed beats accuracy for sustained engagement
Source
How I AI — Claude Duet is beautiful. I still forgot to use it (October 21, 2026) · at 7:03 ▸
When this applies
when choosing between a higher-quality-but-slower option and a faster-but-slightly-lower-quality option for a recurring interactive workflow
The takeaway
For synchronous or near-synchronous collaborative workflows, response speed often outweighs a modest quality advantage.
Why it works
Fast responsiveness keeps the user actively engaged step-by-step like working with a responsive colleague; slowness causes disengagement and task abandonment regardless of eventual quality.
One level up
Judge a tool's fit for a workflow by the interaction cadence it enables, not solely by the quality of its final output.
Where it stops holding
This stops working when the user is working asynchronously and doesn't need to stay engaged in real time, in which case quality may matter more than speed.
Detail
approach: Evaluate an AI colleague's usefulness by whether its latency sustains your working rhythm.
signals:
- agent takes minutes to respond
- user moves to the next task before agent finishes
- user forgets about the task
rationale: A colleague who is slow, even if more capable, gets avoided in favor of a faster, 'good enough' colleague.
In their words
if that colleague is slow, if that colleague goes to a meeting to get work done, like, you kind of forget and don't want to go to that colleague over your more responsive ones, even if the more responsive ones do like 10% worse of a job.
Raise the bar for displacing an incumbent tool
Source
How I AI — Claude Duet is beautiful. I still forgot to use it (October 21, 2026) · at 0:00 ▸
When this applies
when evaluating whether to switch to a new tool that must unseat an already-adopted incumbent in your workflow
The takeaway
When a user already has working solutions, a new product must be exceptional, not merely good, to earn a place in their stack.
Why it works
Switching costs and existing habit loops mean a new entrant must clear a much higher bar than 'good' to earn attention share, because the cost of adopting it is not just quality but also re-training routines.
One level up
Evaluate new tools against the switching cost they create, not just their standalone quality.
Where it stops holding
This stops working when the user has no incumbent solution yet, in which case 'good enough' is sufficient to win adoption.
Detail
rule: A new entrant into a category you already use must be exceptional, not just good or great, to win your attention.
threshold: Must beat the status quo decisively, not incrementally.
rationale: Attention and workflow habits are scarce; incremental improvement doesn't justify the switching cost of learning a new tool.
In their words
to compete for my attention, to compete for being my number one personal agent, you have to be not good, not great, you have to be exceptional.
Re-engagement requires productive annoyance
Source
How I AI — Claude Duet is beautiful. I still forgot to use it (October 21, 2026) · at 4:03 ▸
When this applies
when designing any system (product, email, app) meant to pull a busy user back to re-engage with it
The takeaway
A re-engagement mechanism must be assertive enough to interrupt competing attention, or it will be ignored by design.
Why it works
Subtle, non-intrusive notifications get buried under competing demands on attention; without a disruptive signal, the task silently falls out of the user's working memory.
One level up
Growth and retention mechanics sometimes require trading off 'elegance' for effectiveness.
Where it stops holding
This stops working when users are already highly motivated self-starters who check the tool proactively, or when excessive notifications would cause them to churn from annoyance.
Detail
approach: Judge a proactive agent's notification design by whether it actually resurfaces unfinished work, not by whether it feels tasteful.
signals:
- user forgets tasks were kicked off
- no desktop presence to prompt return
- morning briefings fail to differentiate
rationale: Elegance that doesn't interrupt attention fails to generate behavior change.
In their words
I need you to be a little bit more annoying. I need you to be a little bit more proactive. I need you to be a little bit more noisy or I'm truly going to forget you in the ecosystem of all my agents.
Multiplayer sharing is essential for SMB/team-oriented products
Source
How I AI — Claude Duet is beautiful. I still forgot to use it (October 21, 2026) · at 14:03 ▸
When this applies
when a tool targets a collaborative user segment (small teams, SMBs) but is architected for single-player use only
The takeaway
A product's architecture (single-player vs multiplayer) must match the actual working structure of its target customer segment, not just their job title.
Why it works
SMB owners and busy professionals rarely work in isolation; without sharing/collaboration features, the tool's value is capped to what one person can do alone, undercutting its stated target market.
One level up
Validate that a product's core interaction model fits the real unit of work (individual vs team) for its claimed target customer, not just the persona label.
Where it stops holding
This stops working when the target users genuinely work solo despite nominally being 'business owners' (e.g., true solopreneurs with no staff).
Detail
rule: If a tool targeted at teams/SMBs can't be shared with the user's actual team, it's unlikely to become part of their working stack.
threshold: Needs multiplayer sharing (not just channel-like UI) to fit team workflows.
rationale: Channels that resemble Slack but lack sharing don't replicate Slack's team-effectiveness value.
In their words
If I don't have a way to share my agents with them, I'm not sure if it's gonna worm its way into my working stack.
Slow ticket-shopping task breaks engagement loop
Source
How I AI — Claude Duet is beautiful. I still forgot to use it (October 21, 2026) · at 7:03 ▸
When this applies
when a user delegates a task to an async agent expecting quick turnaround, but the agent takes many minutes to respond
The takeaway
A slower but more capable agent loses out to a faster, slightly less capable one because speed sustains the user's working attention loop.
Why it works
Long task latency means the user's attention has moved to something else by the time the agent finishes, so the output goes unnoticed and the tool is deprioritized even if the work was good.
One level up
In human-in-the-loop systems, response latency often matters more for adoption than raw output quality.
Where it stops holding
This stops working when the task is inherently long-running and the user explicitly expects asynchronous, delayed completion (e.g., deep research tasks).
Detail
situation: Host asked Claude Duet to find good Warriors tickets while he was driving; the task took 10+ minutes with no noisy follow-up.
decision: Host moved on to other things and forgot about the task because the agent wasn't noisy or fast enough to keep him engaged.
outcome: Even though the task was tracked in a checklist, the slow performance made the agent 'fall to the bottom of useless' for the host.
lesson: Even a highly intelligent model loses to a faster, slightly worse one because speed keeps the user in the loop.
In their words
I kind of forgot that it did because it took so long. It didn't close the loop.