Super User Daily: 2026-05-14
Today's pool skews quality over quantity. One vision experiment pushed Claude Code through 200+ iterations to hit 86.1% on a 10-class ImageNet slice without a single neural net — pure symbolic features, around twenty hours, $200 in API cost. A designer mapped a clean 6-phase POC workflow from Figma to shipped prototype. A Japanese PR pro built a one-click weekly content generator and is half-seriously chasing the 300k yen single-shot meme. And the sharpest warning of the day: Claude Code's auto-generated MEMORY.md persists across exits, which means stale habits or bad learning can quietly hijack the next session if you don't know where to find it.
@learningPikachu [Claude Code]
https://x.com/learningPikachu/status/2054031405038920000
Ran a 20-hour, ~$200 Opus loop where Claude Code iteratively built symbolic features and scoring rules over 200+ eval iterations on a 10-class ImageNet subset. Went from 12.7% to 86.1% top-1, no gradients, no learned weights — every classification fully explainable ("58% golden-brown pixels, warm blob covering 71% of image, nature context 0.71"). When it failed on a brass teapot it could literally say why: "brass and fur are the same color at 64x64." The ceiling at 86% is representation saturation, not algorithm. The kind of result that makes you rethink what "training" even means when an agent can edit its own program for twenty hours straight.
@llsbetdigital [Claude Code]
https://x.com/llsbetdigital/status/2054144205551796693
Documented a concrete 6-phase POC workflow from idea to MVP: Figma screens to define vision → describe to Claude and shape UX → map user flows → split the build plan into phased prompts (not one master prompt) → write a Design.md to lock the visual language → ship in Claude Code with Figma links + screenshots + prompts. The real-leverage insight: it's not the prompting itself, it's translating design intent into a build system the AI can actually execute. Working through 6 implementation phases for a travel-planning app to prove the loop.
@Nagatarotaro [Claude Code]
https://x.com/Nagatarotaro/status/2054019397325631571
A startup PR person at his company is using Claude Code together with Google Apps Script to build her own internal tools when off-the-shelf software doesn't fit. He's calling her "PR (physical/technical)" and recommending the writeup to "every business person." Concrete signal that non-engineers are operating Claude Code at a level where it ships real automations — not demos, but workflows the company actually depends on.
@oragnes [Claude Code]
https://x.com/oragnes/status/2054049072210919663
Recommends two Rust-based tools sitting in front of Claude Code: RTK intercepts bash output (git diff, cargo test, grep) and compresses noise before it hits the model — claims 60-90% context savings, hundred-line compile errors squished to twenty. 9Router adds a three-tier seamless fallback: paid subscription first, then cheap APIs like GLM, then free pools like Kiro/OpenCode, with format translation between OpenAI and Claude. Direct response to the two most common Claude Code pain points: context bloat from terminal noise and rate-limit interruptions mid-session.
@akirahamster [Claude Code]
https://x.com/akirahamster/status/2054099821074776120
Building a 3D Enigma cipher machine simulator with interactive Rotor / Plugboard / Lumpboard parts. Compared Meshy ($900 first-month plan, 800-char prompt cap), Codex, and Claude Code across the 3D pipeline. Concluded Meshy beats Claude Code on texture quality but loses badly on structural decomposition — you can't get clean separable parts out of it. Settled on Codex/Claude Code for the structural and operational design, planning to layer Meshy back in for surface polish once Claude Code's texture quality catches up. Useful cross-tool comparison from someone actually shipping a non-trivial 3D app.
@ganbaru_bonjin [Claude Code]
https://x.com/ganbaru_bonjin/status/2054208158411977159
Built a Claude Code workflow that generates a full week's worth of social posts in one click. Says quality has gone up too — every "fix this" prompt makes the output noticeably better. The "300k yen in one shot" meme (a Japanese productivity meme for high-leverage solo income) is starting to feel realistic. Short tweet, but the workflow itself is the case.
@ya_man_Gd77209 [Claude Code]
https://x.com/ya_man_Gd77209/status/2054133701664239858
Discovered that Claude Code's auto-generated MEMORY.md persists after exit — if Claude picks up a weird habit in one session, the only fix is to manually delete files in ~/.claude/projects/<project>/memory/. Argues this matters less for one-shot coding but a lot for office/admin work where the same routine repeats and silent drift could lead to AI doing something unexpected without the user noticing. Notes that the /memory command shows contents but doesn't reveal the file location. Thinking about building a memory-clearing tool. Real warning that nobody talking about Claude Code productivity is mentioning.
🗣 User Voice
User Voice
Memory transparency is the next sharp edge. @ya_man_Gd77209 captures it cleanly: users want to know what Claude Code remembers across sessions and how to clear it without spelunking through ~/.claude/projects/.
Context bloat from bash is real and people are building external compressors. @oragnes shipped a Rust proxy because hundred-line compile errors and git diffs were eating his context window.
Phased prompts beat the master prompt. @llsbetdigital's six-phase workflow is the second time this month a serious builder has explicitly called out that splitting the plan beats trying to one-shot it.
Non-engineers are now first-class Claude Code users. @Nagatarotaro's PR colleague using Claude Code with GAS is a credible signal that the tool is escaping the IDE.
Creative-pipeline tools still have a gap. @akirahamster says Claude Code wins on structure but loses on texture vs Meshy — implicit ask for the next generation of Claude Code's image/3D quality.
Memory transparency is the next sharp edge. @ya_man_Gd77209 captures it cleanly: users want to know what Claude Code remembers across sessions and how to clear it without spelunking through ~/.claude/projects/.
Context bloat from bash is real and people are building external compressors. @oragnes shipped a Rust proxy because hundred-line compile errors and git diffs were eating his context window.
Phased prompts beat the master prompt. @llsbetdigital's six-phase workflow is the second time this month a serious builder has explicitly called out that splitting the plan beats trying to one-shot it.
Non-engineers are now first-class Claude Code users. @Nagatarotaro's PR colleague using Claude Code with GAS is a credible signal that the tool is escaping the IDE.
Creative-pipeline tools still have a gap. @akirahamster says Claude Code wins on structure but loses on texture vs Meshy — implicit ask for the next generation of Claude Code's image/3D quality.
📡 Eco Products Radar
Eco Products Radar
No single tool crossed the 3-mention threshold today. Worth tracking: Cursor and Codex (2 each, in the same workflows as Claude Code), and the new Rust-based context-compression layer RTK + 9Router which is the most concrete attempt yet to fix Claude Code's bash-output bloat at the system level.
No single tool crossed the 3-mention threshold today. Worth tracking: Cursor and Codex (2 each, in the same workflows as Claude Code), and the new Rust-based context-compression layer RTK + 9Router which is the most concrete attempt yet to fix Claude Code's bash-output bloat at the system level.
Comments