Every time I hit a wall with one model, I used to paste my whole thread into another one and explain the project from scratch. That part wasted more time than the actual work. llmwise.ai puts 16 models from 7 companies in one chat, on one plan. Switch models mid-chat and the next model sees the whole conversation. No re-explaining your project. I pick a model per message, and the message count shows before I send. My usual move: start on a fast model for the boring first pass, then switch to a stronger one for the hard part, same thread. Pro is $20 a month plus tax. The monthly allowance is mixable, up to 125 messages on Claude Sonnet 5-class models, plus 60 fast messages a day. Helps anyone juggling more than one AI subscription.

Connecting with builders working on AI and agents. I am a senior SRE. 10 years with Kubernetes, AWS and Terraform. Now I build AI agents for small teams. Lately I talk to my agent more than I type. Voice coding is faster for me. I want to know who else works this way. I also write arXiv papers on chain-of-thought faithfulness. Reply with what you're building. I follow back builders.

Agents drift. They start on your task. Then they wander into other files, other goals, other plans. These 10 prompts keep an agent on task. Paste them into your system prompt, or send them mid-run when the agent starts to drift. 1. "Restate my task in one sentence. List the steps you will take. Wait for my OK before you start." 2. "Put the task goal at the top of every reply. If the goal changes, stop and tell me." 3. "You have 5 tool calls for this step. Stop at 5. Report what you have so far." 4. "Before each action, name the task step it serves. If it serves no step, skip it." 5. "Touch only these files: <list>. Ask me before you read or write anything else." 6. "Do not rename, reformat, or refactor code outside the target function." 7. "Every 5 steps, compare your work to the original task. Report any drift in 2 lines." 8. "Stop and ask me when a test fails twice, a file is missing, or you need a new dependency." 9. "End every reply with one word only: DONE, BLOCKED, or NEED INPUT." 10. "Keep a decision log. Add one line per decision. Do not rewrite old lines." Save this list. Paste it into your agent config today. Reply with the prompt you would add.

Hello @X I am looking for people who are founders and love indie hacking , love to connect. #connect

Two queue designs look fine on day one. They break in different places. One worker per tenant. Each tenant gets its own queue and its own worker. A noisy tenant cannot starve the others. That is the main benefit. The cost is idle workers. A tenant that sends one job an hour still holds a worker. Tenant count grows, worker count grows, and the bill grows with it. Shared pool. All jobs land in one queue, and a pool of workers pulls from it. Simple to run. Cheap when traffic is low. But one tenant with a big backfill can fill the whole queue. Everyone else waits behind it. This is head of line blocking. Where each one breaks: Per tenant workers break on tenant count. You must create queues at runtime and delete them later. Most teams forget the delete step. Empty queues pile up. Workers sit idle. Shared pools break on one loud tenant. A single retry loop or a backfill job can push small fast jobs behind it for minutes. The fix I use most: a shared pool with per tenant limits and fair scheduling. Round robin across tenants. Cap each tenant at a fixed number of in flight jobs. Checklist before you pick: 1. Measure jobs per tenant per hour. If the spread is small, a shared pool is fine. 2. If the spread is wide, keep the shared pool and add a per tenant concurrency cap. 3. Split long jobs into their own queue. Short jobs stay fast. 4. Make every job idempotent. Retries will happen. 5. Add a dead letter queue. Alert on its size, not on error rate. 6. Delete idle per tenant queues after a timeout. Gotchas: Per tenant queues leak. Shared pools hide one slow tenant until the backlog grows. Autoscaling on queue depth fights fair scheduling, because the scaler adds workers for the loud tenant too. Concrete detail: in Celery, set worker_prefetch_multiplier to 1. The default is 4. With 1, a worker takes one task at a time, so one tenant cannot hoard the pool. In BullMQ, use per queue rate limits and a separate queue for long jobs. Start with the shared pool. Add the per tenant cap on day one, not after the first incident. What does your queue look like in production, and what broke first?

i run my whole stack on subscriptions. auth, hosting, db, email, logs, ai. each one is cheap on its own. together they add up, and each one wants its own login and its own billing page. where is your line? how many paid tools do you keep as one person building?

I stopped making one branch per feature. Now I merge to main every day. Feature flags do the job branches used to do. A branch hides code from production. A flag hides a feature inside production. The branch keeps the code out. The flag lets the code in and keeps it off. For a solo builder who ships daily, the flag is usually the better tool. Here is the loop I use. 1. Cut the branch small. Merge it the same day. A branch that lives for a week turns into a merge conflict. 2. Put the new code behind a flag. The flag starts off. 3. Deploy with the flag off. Check that nothing broke. 4. Turn the flag on for your own account. Use the feature for real work. 5. Turn it on for one more person. Watch the logs. 6. Turn it on for everyone. 7. Delete the flag when the feature is stable. Gotchas I hit: A flag does not fix bad code. It hides bad code. The bug is still there when you turn it on. Two flags on the same code path get confusing fast. Keep one flag per feature. Flags add if statements. Put them at the edge of the system, not deep in the logic. If the flag reads from a config file, a typo can turn it on by accident. Make off the default. Stale flags pile up. A flag with no owner is a hidden branch. Nobody knows what it does. Branches still have a place. Use a branch for a risky refactor or a dependency upgrade. A flag cannot hide a build that does not compile. One concrete rule: name the flag `ff_<feature>`, and give it three values only: `off`, `internal`, `on`. No booleans with extra states. Then `git switch -c feat/invoice-export`, merge the same day, and let the flag hold the release. What is your rule for deleting a flag? Do you remove it the day the feature ships, or do you keep it as a kill switch?

i test a new video tool almost every week. most of them feel the same after a day. the one that stuck for me was a small audio cleanup tool. i ran it on a bad interview recording and it just worked. what video or audio tool actually surprised you this year? not the one with the best launch post, the one you still open.

OpenAI dots are really great as an assistant

I have completed feature which is called connectors in LLMWISE.ai what it does basically it uses @composio connectors Also I built layer on top so even models like DeepSeek , GLM whichever support tools can use those connectors. Please try this out it is really good feature and give me feedback #building #IndieVibes

I have completed feature which is called connectors in LLMWISE.ai what it does basically it uses @composio connectors Also I built layer on top so even models like DeepSeek , GLM whichever support tools can use those connectors. Please try this out it is really good feature and give me feedback

I'm building llmwise.ai in public. Here's why. Every day I asked the same question to Claude, then GPT, then Gemini. Three tabs. Three plans. Copy, paste, repeat. The worst part: hitting a limit mid-task with no warning. So I'm building the tool I wanted: • 17 models in one chat • Switch mid-chat and the next model sees the whole thread • Your message count sits next to the send button • One $20 plan, no 5-hour lockouts SRE by day, so this gets built on nights and weekends. I'll share what works and what breaks right here. What would make you switch from your current AI plan?

Hey X 👋 Friday #connect round for people who ship. Reply with your lane and I'll follow back: • AI Agents / MCP • Claude Code / Codex • SRE / DevOps • SaaS Founder • Evals / AI Safety • Weekend builder Tell me what you're shipping this weekend 🤝

Dots with better WiFi

Finally perplexity has same Jev kind of model as decision api like open api

Anthropic is onto something