Dipak (@doctype5) header
Dipak (@doctype5) profile picture

Driving thought leadership around UX design for B2B and B2C products. Passionate about UX, leadership, and continuous improvements.

Pune, India
Joined September 2009
Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)Photo by Dipak (@doctype5)
Dipak retweeted
Anthropic just published the full prompting guide for Claude Fable 5.2. A new models read the same words differently, so your prompt has to change with it: 1. Tell it the code is yours. Add one line like "This codebase is yours and you'll keep building on it." Without it, the model piles new code on top of the old and only reports problems it sees nearby. With it, it cleaned up repeated code in half the tries instead of under 1 in 10. 2. It quits long tasks too early, especially at low and medium effort. Tell it to keep going and work through ordinary errors and timeouts on its own. That cut unfinished sessions in half. 3. "Answer directly, don't overthink" doesn't work anymore. It thinks anyway. If simple requests are slow, lower the effort level instead. 4. Clean out old notes and memory files. One old "never retry" note dropped retries from almost always to 4 in 10. 5. Say exactly what you approve. "Yes, run the migration on staging" works better than "go". 6. High effort is enough for most work. Almost all the gain is already there, and going higher mostly just burns tokens. 7. Strict rules for its reports backfire. Tell it "max 20 words per sentence" and it follows that so exactly that the summary of what it did turns into choppy fragments nobody can follow. Drop the rules you don't really need. The model got much better at listening. The problem is that most of our prompts were written for models that didn't, so they're full of old warnings it now takes seriously.
27
49
1
676
80,995
Dipak retweeted
How to Build an AI Agent in Just 10 Minutes
3
56
402
13,999
Nobody has ever explained all of Claude to you. But here's everything about Claude you must know: Level 1. Don't be lazy, use the Claude desktop app. You will still use the browser 90% of the time, and that's fine. But the day you start using the Claude desktop app, you'd be able to work with your folders without ever uploading the files one by one. You'd use Claude code more. There is no coming back from there. Level 2. You have better things to do than spending 1 minute choosing the best model. Default should be: Opus 5.5 on high. That's it. Do not switch the models mid-chat. It consumes more tokens because Claude re-reads the entire chat (again) to get the context. Level 3. Either Medium or High effort. Nothing else Medium when it is low-effort work. High for most of the tasks. But Extra High mostly burn through your usage limit. Level 4. Delete your old Claude setup. Delete the “Instructions for Claude” box in settings. Delete the old added Claude skills. Old skills and prompts might be slowing you down since they were written for old models. Instead, give one goal and end prompts with: “Before you start, restate my goal in your own words.” Level 5. Ask for HTML. Not for a document. Add two things to your prompt: “Make it an HTML page” and “for [who will read it].” Claude will code the image, doc, or any other format. There is no scope of making mistakes. Level 6. Protect what you plug in. Never click the thumbs up or down under an answer. You are basically giving feedback to AI, and they are allowed to send it to their training data. On a Team plan, once memory is on, leave it on: if an owner turns it off, every member’s memories are deleted permanently. Your first 20 minutes, in order: 1. Install desktop app. 2. No settings are the best. 3. Connect ONE tool - not everything. 4. Use Claude Code more than you do. 5. Delete the installed skills, and add the new ones (written for the latest models) 6. Go to skillsclau.de for my Claude skill library.
37
70
506
30,247
10 Claude Projects to build this weekend:
17
98
703
78,856
🎯𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗖𝗵𝗲𝗮𝘁 𝗦𝗵𝗲𝗲𝘁 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗶𝘀 𝗺𝗼𝗿𝗲 𝗽𝗼𝘄𝗲𝗿𝗳𝘂𝗹 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝘀𝘁𝗼𝗽 𝘁𝗿𝗲𝗮𝘁𝗶𝗻𝗴 𝗶𝘁 𝗹𝗶𝗸𝗲 𝗷𝘂𝘀𝘁 𝗮𝗻 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗰𝗵𝗮𝘁𝗯𝗼𝘁. The real advantage comes from knowing 𝗵𝗶𝗰𝗵 𝗰𝗼𝗺𝗺𝗮𝗻𝗱 𝘁𝗼 𝘂𝘀𝗲, 𝘄𝗵𝗲𝗻 𝘁𝗼 𝘂𝘀𝗲 𝗶𝘁, 𝗮𝗻𝗱 𝗵𝗼𝘄 𝘁𝗼 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝘆𝗼𝘂𝗿 𝗽𝗿𝗼𝗷𝗲𝗰𝘁 𝗮𝗿𝗼𝘂𝗻𝗱 𝗶𝘁. Here’s a practical Claude Code cheat sheet worth saving 👇 𝗖𝗟𝗔𝗨𝗗𝗘.md Project-specific rulebook Claude reads before every task. /init first to auto-generate. Focus on architecture, coding style, build commands, and review checklists. Keep it under 150 lines. 𝗦𝗸𝗶𝗹𝗹𝘀 Reusable markdown prompt templates for standardized workflows. Encode 𝗰𝗼𝗱𝗲 𝗿𝗲𝘃𝗶𝗲𝘄, 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝘀𝗰𝗮𝗻𝘀, 𝗮𝗻𝗱 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 𝗰𝗵𝗲𝗰𝗸𝗹𝗶𝘀𝘁𝘀 to align team standards. 𝗠𝗖𝗣 Connects Claude to local tools, databases, and external APIs. Link MCP servers to query live databases or terminal logs during debugging. 𝗔𝗴𝗲𝗻𝘁𝘀 Spawns background sub-agents for parallel, isolated tasks. Have 𝗮𝗴𝗲𝗻𝘁𝘀 𝘄𝗿𝗶𝘁𝗲 𝘂𝗻𝗶𝘁 𝘁𝗲𝘀𝘁𝘀 while you focus on core feature logic. 𝗣𝗹𝗮𝗻 Forces an architectural outline before any code is written. Critique the plan first, then command Claude to “implement it.” 𝗖𝗼𝗺𝗽𝗮𝗰𝘁 Compresses history to maximize the context window. Use when Claude feels sluggish to keep the context focused. 𝗔𝘂𝘁𝗼𝗱𝗿𝗲𝗮𝗺 AI “sleep cycle” that organizes and refines project memory. Useful for reducing 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗽𝗼𝗹𝗹𝘂𝘁𝗶𝗼𝗻 in long-running workflows. 𝗥𝗮𝗹𝗽𝗵-𝗹𝗼𝗼𝗽 Autonomous iteration loop for persistent self-correction and testing. Ideal for TDD. Always set safe limits for iterations. --- 𝗣𝗿𝗼 𝗧𝗶𝗽𝘀 𝗬𝗼𝘂 𝗖𝗮𝗻 𝗠𝗶𝘀𝘀 ✅ /init → creates your CLAUDE.md starter instantly ✅ Use /context to check usage before things get heavy ✅ /memory → view and edit auto notes anytime ✅ Hooks can run auto-lint and formatting after edits ✅ Always 𝗿𝗲𝘃𝗶𝗲𝘄 𝗱𝗶𝗳𝗳𝘀 before accepting changes ✅ /plan mode = safer exploration before editing ✅ 𝗦𝗸𝗶𝗹𝗹𝘀 + 𝗮𝗴𝗲𝗻𝘁𝘀 = more consistency across workflows ✅ /compact at ~70% context = a useful trigger before conversations become unwieldy --- 𝗢𝗻𝗲 𝗺𝗼𝗿𝗲 𝘁𝗵𝗶𝗻𝗴 𝗜’𝗱 𝗮𝗱𝗱: Don’t ask Claude Code to “build the feature” and hope for the best. Try this workflow: 𝗣𝗹𝗮𝗻 → 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁 → 𝗧𝗲𝘀𝘁 → 𝗥𝗲𝘃𝗶𝗲𝘄 → 𝗖𝗼𝗺𝗽𝗮𝗰𝘁 → 𝗜𝘁𝗲𝗿𝗮𝘁𝗲 The goal isn’t to give Claude more instructions. The goal is to give it 𝗯𝗲𝘁𝘁𝗲𝗿 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲. That’s where AI coding starts feeling less like autocomplete and more like an engineering workflow. 𝗪𝗵𝗶𝗰𝗵 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗰𝗼𝗺𝗺𝗮𝗻𝗱 𝗱𝗼 𝘆𝗼𝘂 𝘂𝘀𝗲 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁? 👇 #ClaudeCode #AI #Coding #SoftwareEngineering #AITools
2
5
27
1,476
AI doesn’t care about your job title uxdesign.cc/ai-doesnt-care-a…
1
4
6
1,544
The taste paradox: why the skill everyone says matters is the one we’ve stopped building uxdesign.cc/the-taste-parado…
2
8
3,310
AI skill atrophy: Are we forgetting how to think? uxdesign.cc/ai-skill-atrophy…
6
1
14
6,159
Save time with Wispr Flow (4x faster than typing) bit.ly/side-wis1 #ad #UX #productivity #AI
2
3
4,091
How to de-slopify your designs uxdesign.cc/how-to-de-slopif…
2
8
1
43
8,787
Designing for AI means designing like it’s 1999 uxdesign.cc/designing-for-ai…
1
5
5,421
The prompt is not an interface uxdesign.cc/the-prompt-is-no…
4
19
5,416
Dipak retweeted
TOP 50 CLAUDE SKILLS YOU NEED RIGHT NOW
9
28
97
3,331
9 High ROI Skills
44
3
184
6,765
How to lead when you're not in charge
3
150
1
586
14,642
The ultimate marketing cheat sheet
19
90
4,150
Dipak retweeted
Best YouTube Channels to learn AI in 2026: 1. AI Explained 👉 youtube.com/@aiexplained-o… 2. Andrej Karpathy 👉 youtube.com/@AndrejKarpath… 3. Cole Medin 👉 youtube.com/@WesRoth/ 4. DeepLearningAI 👉 youtube.com/@Deeplearninga… 5. Futurepedia 👉 youtube.com/@futurepedia_i… 6. Matthew Berman 👉 youtube.com/@matthew_berma… 7. Skill Leap AI 👉 youtube.com/@SkillLeapAI/f… 8. Tech With Tim 👉 youtube.com/@TechWithTim/v… 9. Tina Huang 👉 youtube.com/@TinaHuang1/vi… 10. Two Minute Papers 👉 youtube.com/@TwoMinutePape… Follow me @Tech_Rose1 for more AI IDEA.
25
104
3
242
10,109
AI won't make every designer 10x—but it is changing how great products get designed bit.ly/uxc-lab10 #ad #UX #design
1
19
5,573