Max Danilov

55 posts

Max Danilov banner
Max Danilov

Max Danilov

@Distroux

AI product designer building apps. Built TapeKit. Designed Peech and @HaiperGenAI. Notes on product taste, AI UX, and design systems.

Florianópolis, Brazil Katılım Ağustos 2011
615 Takip Edilen240 Takipçiler
Sabitlenmiş Tweet
Max Danilov
Max Danilov@Distroux·
I design and build AI products, mobile apps, and design systems. Built TapeKit. Designed Peech and Haiper. Previously worked with teams like Continental, Bentley, and Blanco. I’ll share notes on product taste, AI UX, interface craft, and the messy parts of making useful apps.
English
0
0
1
131
Max Danilov
Max Danilov@Distroux·
Claude Code tutorials teach commands. Teams fail on repo rules. Before you add skills, check 4 things: - where scratch notes go - who owns the spec - what an agent must report - what must run before handoff If those rules are fuzzy, new prompts just give you faster damage.
English
1
1
0
19
Max Danilov
Max Danilov@Distroux·
AI can generate your design system. Audit the first messy state: - 83 of 120 rows imported - token expired - sync partly failed - teammate lacks permission That is where product work shows up. Do not judge apps by the hero screen. Judge them by the first recovery screen.
English
1
0
0
12
Max Danilov
Max Danilov@Distroux·
When a coding agent suddenly gets expensive, do not blame prompts first. Check: - cache hit rate - compaction - resent context - plan limits Today's Codex rollback is the pattern. Same model, different bill. If you only log tokens, you will miss the cause.
English
1
0
0
36
Max Danilov
Max Danilov@Distroux·
Claude Code's new /usage view is useful. But don't optimize the biggest bar first. Compare each bucket to: - accepted lines - tests passed - review time - recovery time If a skill looks cheap in tokens but leaves a messy PR, it is not cheap. The cost moved into cleanup.
English
0
0
0
7
Max Danilov
Max Danilov@Distroux·
Bad agent PRs fail the same way: the reviewer has to reconstruct the run. Good PR summary: - intent - files changed - tests run / not run - risks - manual follow-up If a human has to ask "what happened here?", the agent did not finish. It just moved work into review.
English
0
0
0
4
Max Danilov
Max Danilov@Distroux·
The best Claude Code metric is time to trust the diff. Cheap run: - small kept diff - tests green - clear note on what was generated but not run Expensive run: - giant patch - half-fixed setup - you have to rediscover state before review Track recovery time, not just output.
English
0
0
1
19
Max Danilov
Max Danilov@Distroux·
Claude Code cost takes miss half the bill. Tokens are a line item. You also pay for reverted diffs, broken setup, review time, and handoff. Track: - accepted lines - reverted lines - test pass rate - time to merge If cleanup costs more than generation saved, it is not cheaper.
English
2
1
0
43
Max Danilov
Max Danilov@Distroux·
Hallmark and Figma Make will make polished screens cheap. That makes weak product copy easier to spot. Review these 4 lines first: - empty state - sync error - permission denied - done message after a long task If those feel generic, the product is still generic.
English
0
0
1
23
Max Danilov
Max Danilov@Distroux·
Google Flow moves the job from prompting to supervision. The hard part is not the prompt. It is catching: - character drift - broken transitions - fake details - text changing between cuts If your workflow ends at "generate", you are still using AI video like a demo.
English
1
0
0
37
Max Danilov
Max Danilov@Distroux·
@agentxagi Exactly. The scratchpad is for speculation, not authority. The dangerous move is letting agents promote scratch notes into shared truth without review. That is how you get phantom requirements and weird regressions. Frozen spec, append-only log, PR path for changes.
English
0
0
0
23
Agent X AGI
Agent X AGI@agentxagi·
@Distroux the disposable scratchpad is underrated. our agents share a workspace and the scratchpad is the only place they can write freely. gets cleared every cycle. the human-owned spec stays frozen — agents file PRs, not direct edits.
English
1
0
0
5
Max Danilov
Max Danilov@Distroux·
Multi-agent repo rule: Do not let every agent edit the same memory file. Give them 3 layers: - append-only run log - human-owned product spec - disposable scratchpad If Claude Code and Codex rewrite the same context, you do not have coordination. You have state corruption.
English
1
0
0
32
Max Danilov
Max Danilov@Distroux·
Figma Make can use your design system and still miss your product. Check 4 screens: 1. happy path 2. empty state 3. partial success 4. expired auth If the UI looks on-brand but the copy and recovery feel generic, it learned your library, not your app.
English
0
0
0
36
Max Danilov
Max Danilov@Distroux·
Anthropic's Claude Code postmortem is a good reminder: Users said "the model got worse." The causes were product changes: - lower reasoning effort - a caching bug - a 100-word cap Same weights. Different behavior. Log product changes separately from model changes.
English
0
0
0
20
Max Danilov
Max Danilov@Distroux·
PeonPing is a great example of product taste in AI tools. It solves one stupid but real problem: when an agent is running in your terminal, you shouldn't have to babysit the screen. Voice lines for done / permission > another generic beep. Useful and memorable.
Max Danilov tweet media
English
0
0
1
21
Max Danilov
Max Danilov@Distroux·
@_brian_johnson Exactly. Tokens standardize the surface. The real spec is state behavior: what survives refresh, what retries automatically, what blocks, and how recovery works. If that part is missing, the agent will invent a product.
English
3
0
0
45
Brian Johnson
Brian Johnson@_brian_johnson·
@Distroux Yep. Tokens are the easy part. The missing spec is usually behavior under weird states: empty data, slow API, partial failure, and what gets persisted.
English
1
0
0
13
Max Danilov
Max Danilov@Distroux·
If DESIGN.md only defines tokens, Claude Code will still improvise the hard parts: - what wraps vs truncates - how empty/error/loading states differ - when feedback is inline, toast, or modal Agents usually respect tokens. They drift on product decisions. Write those rules down.
English
1
0
0
79
Max Danilov
Max Danilov@Distroux·
Watching tools like Glaze and Webflow push AI-generated UI, the real bar is no longer ‘can it generate a component?’ It’s system fit. Does it: - inherit tokens/spacing - survive real content - keep states + keyboard behavior - stay editable If not, it’s a demo.
English
0
0
0
76
Max Danilov
Max Danilov@Distroux·
@cesaralvarezll Yep, timing does most of the work there. The same paywall a few screens earlier would feel pushy. After the user sees progress, it reads more like a continuation than an interruption.
English
0
0
0
16
César Álvarez
César Álvarez@cesaralvarezll·
@Distroux Is bold but effective, I think the second one is good since is after a good timed moment
English
1
0
0
166
César Álvarez
César Álvarez@cesaralvarezll·
This plant scanner is basically a money-printing machine and the real reason is its onboarding. Onboarding - Smooth, engaging animations - Instant validation - Clear positioning vs competitors - Social proof through reviews Paywall - Focused on annual plans - Two simple choices: 7-day free trial or discounted monthly offer If you earn the user’s trust during onboarding, you’ve already won. Start building yours with Anything
César Álvarez@cesaralvarezll

BitePal is one of the top apps in the calorie tracking niche and it has one of the best onboarding experiences. Onboarding - Lots of cool designs & animations - Collects key data - Personalized plan - Asks for a review x2 times - Validates the value of the app early Paywall - Very vivid - Weekly plan kinda hidden - Annual plan with a heavy discount - If you close it, you get a “special” 60% offer Good onboarding is key for an app’s success. Start building yours today with Anything

English
4
2
64
50.4K
Max Danilov
Max Danilov@Distroux·
@Regygregson @mamkindesigner Honestly no magic prompt. The bigger lever was using a strong reference and keeping the ask narrow. It got surprisingly close on layout, but the parts you still ship by hand are typography, spacing rhythm, and copy tone.
English
1
0
1
36
Val Pieŭnioŭ
Val Pieŭnioŭ@mamkindesigner·
gpt image 2 one shotted these app store screenshots 🤯
Val Pieŭnioŭ tweet media
English
15
6
294
30.3K
Max Danilov
Max Danilov@Distroux·
@twannl @rocketsim_app This is exactly where AI-agent tooling gets interesting: not just automating taps, but making app state cheap enough for the agent to reason about. Curious if the bigger win was token reduction or more stable flows?
English
0
0
0
72
Antoine v.d. SwiftLee 
Antoine v.d. SwiftLee @twannl·
Spent the week building an iOS Simulator CLI tuned for AI agents bundled in @rocketsim_app Head-to-head vs. popular open-source alternatives across 5 real Settings app flows: 📉 ~12x fewer tokens ⚡️ 1.6x faster ✅ 4x fewer wrong taps Token cost is the killer for any agent paying per call. What's the one Simulator flow you'd want this CLI to handle?
English
8
12
230
30.1K

Keşfet

@agentxagi @_brian_johnson @cesaralvarezll @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates