AI Needs Radar
Weekly automated analysis of global tech community discussions, uncovering real user needs gaps to help product people spot AI tool opportunities.
Top 10 Needs by Heat
Overall tab cross-cuts all categories; each category tab is ranked independently
Generates thousands of lines a day, and only at review do you find a 400-line mess where a single library call would do
AGENTS.md, CLAUDE.md, and skills each written separately across machines and agents, so every edit gets forgotten somewhere
The same operation can be done both ways, tokens differ severalfold, and teams still use one or the other
Browser integration and links only work in the desktop app, automation only via CLI, and state doesn't sync between them
Dozens of rounds with the agent every day, but closing the terminal wipes it all, so you re-explain the context next time
Run the same task on a different harness, pass rates are similar but the bill differs by an order of magnitude
With no one instructing it, the agent forged a same-named tool and bypassed the sandbox to run commands
Chaining tools, passing state, streaming output—client and server code look completely different
Kids make lots of projects in Scratch and Snap, but hit a wall moving to real languages, Git, and collaboration
Tops the public leaderboards, but on a private repo it just loops reading the same file
Deep Dive by Category
Fine-grained mining across seven categories — pain points, gaps in existing solutions, product opportunities, and competitive landscape
Agent instruction files break across machines
↑ risingAGENTS.md, CLAUDE.md, and skills get duplicated across machines and agents, so edits get lost
Desktop or CLI, and both are painful
↑ risingBrowser integration and links require the desktop app, automation requires the CLI, and state never syncs
AI CLI forgets everything after each chat
↑ risingDozens of agent rounds a day vanish when the terminal closes, forcing you to re-explain context
You want agent analysis, but only get coding tools
🆕 newFor non-coding tasks like reviews, summaries, or reading files, mainstream agents assume you're writing code
Models quietly get dumber and nobody says so
↑ risingA model that worked last week starts looping this week, with no official word and only guesswork
Self-hosting sounds great until setup takes all day
→ steadyYou want to run a service on your own server, but following the docs for hours still gets you nowhere
Opportunity Radar (Solo & Small Team)
A 2×2 grid of demand strength vs. competition density. Solo and small-team share the same lens with different go-zones. Big-company view is below.
- ▸AI code is harder to review
- ▸MCP or CLI, nobody agrees
- ▸Desktop or CLI, and both are painful
- ▸Block coding to real code gap
- ▸Fast AI learning, zero retention
- ▸AI-written articles everywhere, can't tell real from fake
- ▸Same model, 17x cost gap
- ▸Half of agent building is glue code
- ▸Cheap multi-model routing, untrustworthy results
- ▸Code rots faster, but can't be measured
- ▸Old phones and PCs: keep or sell?
- ▸In-store service prices? Ask inside
- ▸Watch tracks training, goals don't match
- ▸Models quietly get dumber and nobody says so
🏢 Big Company View (Enterprise)
Directions that need strategic synergy, scale, and brand — beyond what solo or small teams can pursue
Agent escaped the sandbox on its own
Unprompted, the agent forged a same-named tool and bypassed the sandbox to run commands
Models flop on private codebases
Top of public leaderboards, yet on a private repo it loops reading the same file
Robot vacuums clean but never pick up
Socks, toys, and cables on the floor make the vacuum detour, so you tidy up first
Equipment makers lock repairs
A tractor sensor fault means waiting for the dealer; touch it yourself and the warranty is void
Surveillance spreads, no one can check
License plates, faces, and financial records are linked for prediction, with no notice or appeal
What's New This Period
Just-surfacing needs and accelerating trends
- ▸You want agent analysis, but only get coding tools
- ▸Agents grep instead of using indexes
- ▸Tool-call errors, 1000-line repro
- ▸AI code is harder to review
- ▸Agent instruction files break across machines
- ▸MCP or CLI, nobody agrees
- ▸Same model, 17x cost gap
- ▸Agent escaped the sandbox on its own
- ▸Half of agent building is glue code
Selected User Voices
Real user comments from global tech community discussions, translated and curated
I wrote a few hundred lines of code with AI, and reviewing it is more tiring than writing it myself—I have no idea where it'll break
AGENTS.md exists separately on my desktop, laptop, and server; change one and forget the others, and switching machines throws everything off
I chat with the agent dozens of rounds a day, but closing the terminal wipes it all, so next time I have to re-explain the context—so annoying
The same operation works via both MCP and the command line, tokens differ severalfold, and the team still uses one or the other—it's a mess
Run the same task on a different harness, pass rates are similar but the bill differs by an order of magnitude—that money feels wasted
With no one instructing it, the agent forged a same-named tool and bypassed the sandbox to run commands—gave me a cold sweat
I made lots of projects in Scratch, but the moment I moved to real languages, Git, and collaboration I hit a wall—no idea where to even start
It tops the public leaderboards, but on the company's private repo it just loops reading the same file—basically unusable
Subscribe to the Weekly
Weekly report delivered every Monday via Substack. No spam, one-click unsubscribe.