DEV Community: Twisted-Code'r

AUTO WRITES PR FOR YOU - PRDRAFT

Twisted-Code'r — Sat, 18 Apr 2026 07:17:35 +0000

Twisted-Code'r

Apr 11

I got tired of PRs that just said "fix" — so I built a GitHub App that writes the description for you

#showdev #automation #github #productivity

Comments

3 min read

I got tired of PRs that just said "fix" — so I built a GitHub App that writes the description for you

Twisted-Code'r — Sat, 11 Apr 2026 10:54:30 +0000

We've all seen it.

You open a pull request from a teammate — or your past self — and the description says:

"fix bug"

That's it. No context. No what changed. No why. You're now a code archaeologist trying to reverse-engineer intent from a diff.

I've been that person. I've also been on the reviewing end, nudging people repeatedly for context that should've been in the description from the start. It's death by a thousand tiny frustrations.

The real problem isn't laziness

Developers don't skip PR descriptions because they don't care. They skip them because by the time the code is done, switching context to write documentation feels like starting a second job. You've been deep in the problem for hours. Writing "what changed and why" in plain English requires a different gear entirely.

So most people don't. Or they write something just good enough to pass review, which ends up useless to anyone reading it six months later.

What I built

I'm a solo developer in Kerala, India, building PRDraft — a GitHub App that reads your PR diff and auto-writes a structured description. Two-click install, zero config.

When you open a PR, it:

Fetches the diff
Sends it to an LLM (Groq / llama-3.3-70b)
Posts a structured description directly to your PR body

The output looks like this:

## What changed
Refactored the auth middleware to use JWT validation instead of session cookies.
Removed the legacy session store dependency.

## Why
Session-based auth was causing intermittent failures in the load-balanced setup.
JWT tokens eliminate the shared state problem entirely.

## How to test
- Log in with an existing account — confirm token is returned in response header
- Test expired token path — confirm 401 is returned correctly

Not perfect every time. But dramatically better than "fix auth" — and it takes zero effort from the developer.

Stack (all free)

Next.js on Vercel
Supabase (PostgreSQL)
Groq API (free tier — 14,400 req/day)
Paddle for payments
GitHub App webhooks via Octokit

Total infra cost: $0/month. Built this in evenings over 4 weeks while learning most of it as I went.

The thing nobody tells you about building GitHub Apps

The hardest part wasn't the AI integration. It was webhook deduplication.

GitHub retries webhooks aggressively if your endpoint doesn't respond in time. I once had a single PR trigger 38 webhook calls. My PR count was wildly inflated, users were hitting the free tier cap immediately, and I had no idea why.

Fix: unique constraint on (installation_id, pr_number) in Supabase, plus a dedup guard before processing. The webhook still fires 38 times — but only the first one does anything.

If you're building on GitHub webhooks, add this early. Don't find out the hard way like I did.

Where it is now

Live at github.com/apps/prdraft
Free for first 5 PRs, then $9/month
2 installs so far (one of which is me)
GitHub Marketplace listing submitted, waiting on review

No paid users yet. That's the next problem to solve.

Try it

If you're tired of writing PR descriptions — or tired of reviewing PRs with no context — install it and let me know what the output looks like on your actual diffs.

Feedback at this stage means everything. I read every reply.

Install PRDraft — free for 5 PRs

Review

Building in public. Day 10 of however long this takes.

I built a GitHub App that auto-writes PR descriptions — here's what happened in 8 days

Twisted-Code'r — Mon, 06 Apr 2026 17:30:22 +0000

I hate writing PR descriptions.

Not because I'm lazy — I know what I changed. Articulating it clearly, in a format that helps reviewers, takes 10 extra minutes of context switching. Every. Single. Time.

So I built PRDraft — a GitHub App that reads your diff when you open a PR and posts a structured description automatically. 2-click install, zero config, no CLI.

Here's what 8 days of building looks like from scratch, solo, from Kerala, India.

Day 1–2: Does anyone actually care?

Before writing a single line of code I posted in GitHub Discussions and dropped a question on Dev.to:

"Who actually writes good PR descriptions consistently?"

Got real responses. Two that stuck:

"Yes, real pain point especially on fast-moving teams"
"I'd trust it if it accurately reflects the diff and I can edit before submitting"

That was enough. Building justified.

Day 3–5: Building the thing

Tech stack I landed on:

Next.js (App Router) + Tailwind
Supabase (PostgreSQL) for installs and PR event tracking
Groq API (llama-3.3-70b-versatile) for AI generation — genuinely free, 14,400 requests/day, works in India
Octokit for GitHub App integration
Vercel for hosting

Why Groq instead of Claude or Gemini:

Anthropic API costs money, no free tier
Google Gemini free tier has limit:0 quota for India region — confirmed bug
Groq is genuinely free and fast

Total infra cost: $0/month

The core flow:

User opens PR → GitHub sends webhook → verify signature → 
fetch diff → send to Groq → post structured description back to PR

The generated description looks like this:

## What changed
- Added webhook signature verification using Octokit
- Introduced free tier cap logic (5 PRs/month) checked before incrementing

## Why
Prevents unauthorized webhook calls and limits free usage before billing is set up.

## How to test
- Open a PR on a repo where PRDraft is installed
- Check that the description is auto-populated within seconds

## Notes
Free tier cap is checked BEFORE incrementing pr_count to avoid off-by-one errors

Day 5: First external user

roshhellwett installed it, opened a PR, got a description within seconds.

Their feedback: "Works as expected — the structure is solid and actually saves time. Groq is noticeably fast. For complex PRs with scattered changes it can feel a bit generic — but as a starting point to edit from? Definitely useful."

That was the moment this stopped being a side project and started being a product.

Day 6–7: Bug fixes and distribution

Bugs I fixed:

Free tier cap was incrementing BEFORE the check — off-by-one error meant users got 6 PRs, not 5
account_login was being overwritten as 'unknown' on every PR event — upsert was hitting the wrong fields
GitHub retries webhooks every 15s on timeout — PR #15 fired 38 times. Fixed with a unique constraint on (installation_id, pr_number)

Distribution channels I tried:

Dev.to — fully open, post live ✅
Indie Hackers — post live, daily updates ✅
GitHub Marketplace — submitted for review ✅
Reddit — karma 3, blocked until Day 9
Hacker News — karma -8, commenting daily to recover

Painful lesson: Every platform blocks new accounts. Don't assume channels are open.

Day 7: Dashboard

Built a full dashboard at /dashboard:

Plan badge (Free / Pro)
PRs used vs cap with a colour-coded progress bar
Recent PR activity — repo, PR number, timestamp
Upgrade CTA
Installation info

GitHub automatically redirects users to the dashboard after install via the Setup URL — no OAuth needed.

Day 8: More dashboard improvements

Shipped today:

Recent activity now shows PR titles alongside repo and PR number
Every generated PR description now has a dashboard link in the footer — so anyone on the team can check usage without hunting for the URL
Free tier cap now posts a comment directly on the PR when the limit is hit, instead of silently doing nothing

The payment provider saga

This is the part nobody talks about.

Building the product took days. Setting up payments is taking longer.

Stripe — invite-only in India since 2024
Lemon Squeezy — international payments blocked for individuals in India
Gumroad — payout setup issues for Indian bank accounts
Paddle — working on it

The product is ready to charge. The money pipe is the blocker.

Update Apr 18: Automated outreach system now running daily. 57+ emails sent to team leads.

Current numbers (Day 8)

Metric	Value
Installs	2
Paid users	0
MRR	$0
PRs described	13
Infra cost	$0/month

What's next

Get payment provider working (Paddle)
Invite more beta users
Reddit posts unlock ~Day 9
HN karma grind back to positive
Month 2: first paid conversion push

Try it

2-click install, no config, free for 5 PRs/month:

github.com/apps/prdraft

Landing page: prdraft.carrd.co

Related: I stopped writing PR descriptions — here's what I did instead

Building in public from Kerala, India. Following along on Indie Hackers → search PRDraft.

I stopped writing PR descriptions. Here's what I did instead.

Twisted-Code'r — Thu, 02 Apr 2026 15:58:14 +0000

I opened 6 PRs last week.
I wrote a good description for exactly one of them.

The rest looked like this:

"fix bug"
"update styles"
"wip"
"changes" (my personal low point)

And I'm not even ashamed — I've seen senior engineers with
10 years of experience do the same thing.

Here's why it happens:

You just finished 3 hours of deep work. You know exactly what
you changed and why. Writing it down feels like explaining a
joke — the effort of articulating it kills the momentum. So
you type something vague and hit submit.

The reviewer then has to reverse-engineer your intent from
the diff. They leave a comment asking what you changed.
You explain it in the comment. Everyone wastes 20 minutes
doing what the PR description should have done in the first place.

So I started building a GitHub App to fix this for myself.

It hooks into GitHub, reads the diff when you open a PR,
and writes the description for you. Structured — what changed,
why, how to test it, what to watch for. Takes about 4 seconds.

I'm still building it. Calling it PRDraft.

Honest question for this community: is this a real pain
for you, or have you just accepted it as part of the job?

If you want it : https://prdraft.carrd.co

I built an open-source tool that stops personal data from leaking into AI chatbots

Twisted-Code'r — Sun, 29 Mar 2026 05:40:46 +0000

Ever copy-pasted something into ChatGPT and immediately
thought "wait, should I have done that?"

If you're building an AI app that handles user data, you need to know what's leaking into your LLM API before a regulator does.

That's the problem ShadowAudit solves.

It sits between your app and any LLM API and scans every
prompt before it leaves your system — catching emails,
phone numbers, API keys, and Indian national IDs like
Aadhaar and PAN numbers.

Two lines to integrate:

sa = ShadowAudit.from_config("shadowaudit.yaml")
client = sa.wrap(openai.OpenAI())

That's it. Everything else stays the same.

It also generates GDPR Article 30 compliance reports
automatically from your audit log — one command, done.

Built this over summer as part of my open-source portfolio.
Would love feedback from the community.

GitHub: github.com/Jeffrin-dev/ShadowAudit

CloudDev v0.3.0 — I built a free LocalStack alternative in Go with 17 AWS services published: true

Twisted-Code'r — Tue, 24 Mar 2026 09:47:56 +0000

If you've ever used LocalStack and balked at the $35/month price tag for the pro version, CloudDev might be exactly what you're looking for.

CloudDev is a free, open-source local AWS emulator built in Go — a single binary, zero runtime dependencies, works with your existing AWS CLI and SDKs out of the box.

Today I'm releasing v0.3.0, which adds 8 new AWS services and brings the total to 17 emulated services.

👉 github.com/Jeffrin-dev/CloudDev

Why I Built This

The "vendor kills free tier, community builds alternative" pattern keeps playing out everywhere. LocalStack's free tier is increasingly limited, and most developers just need basic local emulation for dev and CI — not a $35/month SaaS subscription.

The goal was simple: a tool that:

Works with zero config changes to your existing AWS CLI setup
Ships as a single Go binary with no runtime dependencies
Covers the AWS services that 90% of applications actually use
Is completely free and open source

What's in v0.3.0

🔐 Auth & Security

IAM (port 4593) — users, roles, policies, AssumeRole
STS (port 4592) — GetCallerIdentity, AssumeRole, GetSessionToken
KMS (port 4599) — CreateKey, Encrypt, Decrypt, GenerateDataKey

🏗️ Orchestration

CloudFormation (port 4581) — CreateStack, UpdateStack, parses YAML/JSON templates
Step Functions (port 4585) — state machines, executions, RUNNING → SUCCEEDED transitions

🌉 Events & Cache

EventBridge (port 4587) — event buses, rules, PutEvents with Lambda/SQS routing
ElastiCache (ports 4597/4598) — Redis-compatible TCP server + HTTP management API
Cognito (port 4596) — user pools, users, InitiateAuth with mock tokens

Quick Start

git clone https://github.com/Jeffrin-dev/CloudDev.git
cd CloudDev
go build -o clouddev .

./clouddev init my-app
cd my-app
../clouddev up

That's it. All 17 services start immediately.

Works With Your Existing AWS CLI

No special tooling needed. Just point --endpoint-url at CloudDev:

# KMS — create a key and encrypt/decrypt
aws --endpoint-url=http://localhost:4599 kms create-key \
    --description "my-key" --region us-east-1

aws --endpoint-url=http://localhost:4599 kms encrypt \
    --key-id key-1 --plaintext "hello-world" --region us-east-1

# Cognito — create a user pool and authenticate
aws --endpoint-url=http://localhost:4596 cognito-idp create-user-pool \
    --pool-name my-pool --region us-east-1

aws --endpoint-url=http://localhost:4596 cognito-idp initiate-auth \
    --auth-flow USER_PASSWORD_AUTH \
    --auth-parameters USERNAME=user,PASSWORD=pass \
    --client-id client-1 --region us-east-1

# Step Functions — create and run a state machine
aws --endpoint-url=http://localhost:4585 stepfunctions create-state-machine \
    --name my-machine \
    --definition '{"StartAt":"Hello","States":{"Hello":{"Type":"Pass","End":true}}}' \
    --role-arn arn:aws:iam::000000000000:role/role \
    --region us-east-1

The Web Dashboard

Every service gets a real-time status indicator in the dashboard at localhost:4580. It auto-refreshes every 5 seconds so you can see at a glance what's running.

All 17 services show up with their port and a green/red status dot.

How It Was Built

This project was built with an unusual workflow: Claude as the instructor and Codex as the developer.

Claude designed the architecture, wrote the detailed implementation prompts for each module, debugged issues from test output, and caught protocol mismatches (like the AWS CLI sending AWSStepFunctions.* instead of AmazonStates.* as the X-Amz-Target prefix — something you only discover by sniffing the actual HTTP request).

Codex implemented each module from those prompts, one PR at a time.

It's been a genuinely interesting experiment in human-AI-AI collaboration. The hardest bugs were always the subtle protocol differences between what the AWS documentation says and what the AWS CLI actually sends.

Complete Service & Port Reference

Service	Port
S3	4566
DynamoDB	4569
Lambda	4574
SQS	4576
API Gateway	4572
SNS	4575
Secrets Manager	4584
CloudWatch Logs	4586
IAM	4593
STS	4592
KMS	4599
CloudFormation	4581
Step Functions	4585
EventBridge	4587
ElastiCache (Redis)	4598
ElastiCache (HTTP)	4597
Cognito	4596

What's Next

v0.4.0 will focus on:

CloudWatch Metrics
SQS FIFO queues
Lambda layers
X-Ray tracing

Get Involved

The project is Apache 2.0 licensed and contributions are very welcome. Check out CONTRIBUTING.md to get started.

If CloudDev saves you time or money, a ⭐ on GitHub goes a long way — it helps other developers find the project.

👉 github.com/Jeffrin-dev/CloudDev

I Built a Free Voice-Enabled AI Agent That Uses Notion as Its Brain — Notion MCP Challenge

Twisted-Code'r — Sat, 21 Mar 2026 12:42:40 +0000

What I Built

NotionMind is a free, voice-enabled AI agent that uses your Notion workspace as its long-term memory. You can save notes by speaking, ask questions about your past work, search by meaning not keywords, trace connections across a knowledge graph, and view a stunning terminal analytics dashboard — all at zero cost.

The tagline: Your Notion workspace, but it thinks back.

The Problem I Was Solving

I had hundreds of Notion notes but couldn't actually use them. Searching by keyword misses context. I'd forget I already researched something. I couldn't see patterns in what I was learning. Notes just sat there — static, passive, dumb.

I wanted a brain that:

Remembers everything I tell it
Answers questions from my own notes
Finds connections I missed
Works offline, costs nothing

The Free Stack (Genuinely $0/Month)

This was a hard constraint I set myself — no paid APIs, no credits, nothing that expires.

Component	Tool
AI Brain	Groq — Llama 3.3 70B + Llama 3.1 8B (free tier)
Workspace	Notion API (free)
Semantic Search	fastembed BAAI/bge-small-en-v1.5 (runs locally)
Web Search	DuckDuckGo ddgs (no key needed)
Voice Input	Google Speech Recognition (free)
Voice Output	Microsoft Edge TTS — Jenny Neural (free)
Telegram Bot	python-telegram-bot (free)

Total: $0/month. Not "free tier with limits you'll hit." Actually free.

How It Works — The MCP Architecture

NotionMind uses the Notion MCP to give Groq structured access to your workspace. The agent has three tools:

mcp_search_notes — keyword search across the database
mcp_create_note — create new notes with AI-generated title + tags
mcp_list_all_notes — retrieve recent notes sorted by date

When you ask a question, the MCP agent calls these tools in a loop until it has enough context to answer.

Feature Walkthrough

1. Smart Save

Type or speak anything. Groq auto-generates the title, tags, and date:

>: i fixed the JWT expiry bug in the auth module today
→ routed to: save
✓ Saved!
Title: Fixed JWT Expiry Bug
Tags:  auth, bug-fix, JWT
Date:  2026-03-24

2. Natural Language Routing

This was one of my favorite features to build. Instead of making users type exact commands, Groq reads the input and routes it automatically:

>: who is Virat Kohli
→ routed to: ask

>: show me my recent notes
→ routed to: list

>: remind me to study at 6pm
→ routed to: remind

Exact command names bypass Groq entirely — zero latency for power users.

3. Semantic Search

This is where it gets interesting. Regular search matches keywords. Semantic search matches meaning.

>: graph → 7. Search
What are you looking for: cricketer from Kerala

It finds the note about Sanju Samson even though the search phrase "cricketer from Kerala" doesn't appear verbatim anywhere. The fastembed model converts both the query and all note content into vectors, then measures cosine similarity.

The model runs entirely on CPU — no API call, no internet needed after the first download.

4. Think — Multi-hop Reasoning

Ask a complex question. NotionMind finds relevant notes, expands one hop through the knowledge graph, then asks Groq to reason across all of it:

>: graph → 8. Think
Your question: what have I been researching lately?

It extracts keywords from your question, searches each separately, merges results, pulls in graph-connected notes, and gives you a synthesized answer with exact note citations.

5. Recall — Knowledge Evolution

Pick any topic and see how your understanding evolved over time:

>: graph → 9. Recall
Topic to recall: notionmind

Groq narrates the arc: initial curiosity → experiments → insights → where it seems headed. It references your actual notes and dates.

6. The Dashboard

This is the most visually striking part of the project.

>: dashboard

You get:

GitHub-style 30-day activity heatmap using ASCII heat characters (· ▪ ▫ ▬ █)
14-day knowledge growth bar chart
Topic velocity — tags this week vs last week with ↑ ↓ → trend arrows
Quick stats panel — streak, total notes, top tag, avg per day
Upcoming reminders panel
Pending todos with priority icons

7. AI Insights

>: insights

Four beautifully styled panels:

🔥 Peak Productivity — when you're most active, specific dates cited
🕳 Knowledge Gaps — what you're underexploring given your current work
📉 Fading Topics — what you explored but drifted from
⚡ This Week's Action — one sharp, specific recommendation

8. Knowledge Graph

>: graph → 1. Build

Groq analyses all your notes in pairs and finds genuine content connections — not just temporal ones. The graph is stored locally and built incrementally, so it only checks new pairs on each rebuild.

From there you can:

View an ASCII tree of connections
Find the shortest path between two notes
See all neighbours of a note
Manually relate two notes

9. Suggest — Auto Related Notes

Every time you save a note, NotionMind automatically suggests related ones:

✓ Saved! Title: Watched India vs Australia

Related notes in your brain:
  → research about Virat Kohli (2026-03-22)
  → research about god of cricket (2026-03-23)

The suggestion engine extracts the single main keyword from your note (ignoring noise like "today", "watched", "done"), then matches it against existing note summaries. Fast and surprisingly accurate.

10. Full Telegram Bot

Every feature is available on your phone:

/save   /ask    /today   /list    /search
/semantic       /read    /delete  /stats
/todos  /addtodo        /donetodo
/remind /reminders
/think  /recall  /dashboard      /insights
/weekly /inbox   /results
/menu   /export

11. Todo List

>: todo

Todos stored directly in your Notion database, tagged todo. Add with priority (🔴 high / 🟡 medium / 🟢 low), due date, and notes. Overdue items highlighted in red. Completed items archived cleanly.

12. Voice Mode

>: voice

Speak your note. Google STT transcribes it, you confirm, NotionMind saves it. Responses are read back using Edge TTS (Jenny neural voice online, espeak offline). Six languages supported: English, Hindi, Spanish, French, Tamil, Malayalam.

The Hardest Parts to Build

Semantic search that doesn't hallucinate

My first attempt used Groq to rank notes by relevance. It hallucinated constantly — returning completely unrelated notes with score 1.0. The fix was switching to fastembed for actual vector similarity, which never hallucinates because it's pure math. The tricky part was getting "cricketer from Kerala" to correctly find the Sanju Samson note — the solution was fetching full page block content instead of just the summary field.

The suggest feature

The naive approach (semantic search the full note text) was too slow and too noisy. "I watched cricket match today" would match notes about "daily summaries" because the word "today" was everywhere. The fix was asking Groq to extract a single main keyword first — so "cricket match" becomes just cricket, which then matches only cricket notes.

Command routing strictness

Early versions of the Groq router would confidently route "who is Sachin Tendulkar" to list instead of ask. The fix was a strict JSON schema, an explicit confidence threshold, and specific rules for question words (who/what/when/where/why/how → always ask).

What I'd Add Next

Browser extension to save highlighted web content directly to NotionMind
Daily digest email — Groq summarises yesterday's notes and emails it to you
Obsidian sync — two-way sync with Obsidian vault

Try It Yourself

git clone https://github.com/Jeffrin-dev/NotionMind.git
cd NotionMind
python3 -m venv venv
source venv/bin/activate
pip install notion-client==2.2.1 groq python-dotenv rich \
            httpx ddgs SpeechRecognition edge-tts pygame \
            schedule python-telegram-bot fastembed numpy
cp .env.example .env
# fill in your Notion + Groq API keys
python3 notionmind.py

You need a free Notion account and a free Groq account. That's it. No credit card anywhere.

Built for the Notion MCP Challenge · March 2026 · by Jeffrin

I Ditched GitHub Copilot and Now I Code With AI for Free — Locally

Twisted-Code'r — Sat, 21 Mar 2026 04:25:04 +0000

I was paying $10/month for GitHub Copilot. It's fine. It works. But it means every keystroke I type goes to Microsoft's servers, my code context gets shipped off somewhere, and I'm locked into whatever pricing they decide next year.

Then I found out I could run a better-than-Copilot setup on my own machine, completely free, with no data leaving my computer.

Here's exactly how I did it, and how you can too.

What You Actually Need

Before anything else — honest expectations:

A machine with at least 8GB RAM (16GB is better)
~5GB free disk space per model
A decent CPU, or an NVIDIA/Apple Silicon GPU for speed
About 20 minutes of setup time

No API keys. No credit cards. No subscriptions.

Step 1: Install Ollama

Ollama is the piece that makes all of this possible. It's basically a runtime that lets you pull and run open-source LLMs the same way Docker lets you run containers.

macOS / Linux:

curl -fsSL https://ollama.com/install.sh | sh

Windows:
Download the installer from ollama.com.

Verify it's running:

ollama --version

Step 2: Pull a Coding Model

This is the part where you choose your AI. For coding specifically, these are the ones worth using:

For most people (8GB RAM):

ollama pull qwen2.5-coder:7b

Qwen 2.5 Coder from Alibaba is genuinely impressive. Beats older Copilot versions on HumanEval benchmarks. Specialised entirely for code.

If you have 16GB+ RAM:

ollama pull qwen2.5-coder:14b

Noticeably better at multi-file context and explaining complex logic.

If you're on Apple Silicon (M1/M2/M3):

ollama pull deepseek-coder-v2:16b

Runs fast on Metal. Great at refactoring and docstring generation.

Absolute minimum (4GB RAM):

ollama pull qwen2.5-coder:3b

Smaller but still surprisingly capable for autocomplete and simple functions.

Test it immediately:

ollama run qwen2.5-coder:7b "write a Python function to flatten a nested list"

If you get a clean response, you're good.

Step 3: Pick Your Editor Integration

Now the fun part — making it feel like Copilot inside your actual editor.

VS Code → Continue

Continue is the open-source Copilot alternative. It's a VS Code (and JetBrains) extension that hooks directly into your local Ollama instance.

Install the Continue extension from the VS Code marketplace
Open Continue's config file (~/.continue/config.json) and add:

{
  "models": [
    {
      "title": "Qwen Coder 7B",
      "provider": "ollama",
      "model": "qwen2.5-coder:7b"
    }
  ],
  "tabAutocompleteModel": {
    "title": "Qwen Coder 7B",
    "provider": "ollama",
    "model": "qwen2.5-coder:7b"
  }
}

Press Tab to accept inline completions, Cmd+I (Mac) or Ctrl+I (Windows) to open the chat panel.

That's it. You now have inline autocomplete, a chat panel, and codebase-aware Q&A — all local.

Neovim → gen.nvim or avante.nvim

If you're a Neovim user, add this to your config with gen.nvim:

require('gen').setup({
  model = "qwen2.5-coder:7b",
  host = "localhost",
  port = "11434",
})

Then :Gen opens a prompt. Select code visually and run :Gen Enhance_Code or :Gen Add_Tests.

JetBrains → Continue (same plugin, different install)

Install Continue from the JetBrains marketplace. Same config file works.

Step 4: Supercharge It With Open WebUI (Optional but Worth It)

Open WebUI gives you a ChatGPT-like interface for your local models. Useful when you want to have a longer conversation about architecture, paste in a whole file, or explain a bug.

docker run -d \
  -p 3000:8080 \
  --add-host=host.docker.internal:host-gateway \
  -v open-webui:/app/backend/data \
  --name open-webui \
  ghcr.io/open-webui/open-webui:main

Open http://localhost:3000, connect to your Ollama instance, and you have a full ChatGPT-style interface running entirely offline.

Real-World Performance

After a month of daily use on a MacBook Pro M2 with 16GB RAM, here's what I found:

Task	Qwen 2.5 Coder 7B	GitHub Copilot
Simple function completion	✅ Excellent	✅ Excellent
Refactoring a 100-line file	✅ Good	✅ Good
Explaining unfamiliar code	✅ Very good	✅ Very good
Multi-file context	⚠️ Limited	✅ Better
Speed (M2 Mac)	~2–3 tok/sec	Near instant
Privacy	✅ 100% local	❌ Sent to servers
Cost	✅ Free	❌ $10/month

Speed is the real tradeoff. On CPU-only machines, responses are slower than a cloud API. On Apple Silicon or an NVIDIA GPU, the gap closes a lot.

The Part Nobody Tells You

Prompting matters more locally than with cloud models.

Cloud models like GPT-4 or Claude have been fine-tuned to be forgiving — they infer what you meant even if you're vague. Smaller local models are more literal. A vague prompt gets a vague answer.

Instead of:

fix this function

Try:

This Python function is supposed to parse a JWT token and return the 
payload as a dict. It currently throws a KeyError when the token is 
expired. Fix the expiry handling and add a try/except that returns None 
on any decode failure.

More context = dramatically better output. Once I adjusted my prompting habit, the quality difference between local and cloud shrank a lot.

Bonus: Free Cloud Options When Local Isn't Enough

Sometimes you need a bigger model for a hard problem. These are genuinely free with no credit card:

Groq — Llama 3.1 70B running at insane speed. Free tier is generous.
Google AI Studio — Gemini 1.5 Flash, 1M token context window, free.
Cerebras — 1M tokens/day free, fastest inference available right now.

You can configure all of these in Continue the same way as Ollama — just swap the provider and add an API key.

TL;DR

# 1. Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# 2. Pull a coding model
ollama pull qwen2.5-coder:7b

# 3. Install Continue extension in VS Code

# 4. Start coding for free

Your code stays on your machine. You pay nothing. It's genuinely good enough for daily use.

The setup takes 20 minutes and you'll never think about it again.

If this helped, I'm posting more practical AI dev workflow stuff — follow along. And if your local setup is different from mine, drop it in the comments — curious what models people are running.

CloudDev v0.2.0 — SNS, Secrets Manager, CloudWatch Logs and more!

Twisted-Code'r — Fri, 20 Mar 2026 04:34:43 +0000

# CloudDev v0.2.0 is here! 🎉

Following the success of v0.1.0, I've released CloudDev v0.2.0
with 8 new modules!

What's new?

3 New AWS Services

📣 SNS — pub/sub with SQS delivery
🔐 Secrets Manager — local secret storage
📊 CloudWatch Logs — log groups and streams

5 New Features

⚡ Lambda Hot Reload — drop a zip, auto-loads in 2 seconds
📋 Status Command — clouddev status
💾 Data Persistence — survives restarts
💰 Cost Estimator — clouddev estimate
☸️ Kubernetes Support — detect services from k8s manifests

Quick Start

clouddev init my-app
cd my-app
clouddev up

👉 GitHub: https://github.com/Jeffrin-dev/CloudDev

If you find it useful please ⭐ the repo!

I built a free, open-source alternative to LocalStack — CloudDev

Twisted-Code'r — Wed, 18 Mar 2026 05:39:19 +0000

I built a free, open-source alternative to LocalStack — CloudDev

If you've been following the news, LocalStack recently ended its
free Community Edition. Developers now need an account and are
being pushed toward paid plans.

So I built CloudDev — a lightweight, 100% free, open-source
local AWS development environment. No account. No paywalls. Forever.

What is CloudDev?

CloudDev lets you run core AWS services locally on your machine
with a single command:

clouddev up

It starts local emulators for:

🪣 Amazon S3 — object storage
🗄️ Amazon DynamoDB — NoSQL database
⚡ AWS Lambda — serverless functions
📨 Amazon SQS — message queues
🌐 Amazon API Gateway — HTTP routing to Lambda
🖥️ Web Dashboard — live service status
🔍 IaC Detection — auto-configure from Terraform/CloudFormation

Why I built it

Developing on AWS is painful:

You have to deploy to the cloud just to test
Provisioning Lambda or API Gateway takes minutes
Accidental usage generates unexpected charges
Hard to test locally without internet

CloudDev solves all of this.

How it works

1. Initialize a project

clouddev init my-app
cd my-app

This creates:

my-app/
  clouddev.yml
  functions/
  infrastructure/

2. Configure services (clouddev.yml)

services:
  s3: true
  dynamodb: true
  lambda: true
  sqs: false
  api_gateway: true

3. Start your local cloud

clouddev up

S3 server starting on port 4566
DynamoDB server starting on port 4569
Lambda server starting on port 4574
API Gateway starting on port 4572
Dashboard running at http://localhost:4580
CloudDev is running. Press Ctrl+C to stop...

4. Use with the real AWS CLI

# S3 — create bucket and upload file
aws --endpoint-url=http://localhost:4566 s3 mb s3://my-bucket
aws --endpoint-url=http://localhost:4566 s3 cp file.txt s3://my-bucket/

# DynamoDB — create table and put item
aws --endpoint-url=http://localhost:4569 dynamodb create-table \
  --table-name Users \
  --attribute-definitions AttributeName=id,AttributeType=S \
  --key-schema AttributeName=id,KeyType=HASH \
  --billing-mode PAY_PER_REQUEST

# Lambda — create and invoke function
aws --endpoint-url=http://localhost:4574 lambda create-function \
  --function-name hello \
  --runtime python3.9 \
  --handler hello.handler \
  --role arn:aws:iam::000000000000:role/test \
  --zip-file fileb://hello.zip

aws --endpoint-url=http://localhost:4574 lambda invoke \
  --function-name hello \
  --payload '{"key": "value"}' \
  output.txt

5. Auto-detect from Terraform

Have an existing Terraform project? Just run:

clouddev detect

CloudDev scans your .tf files and automatically enables
the right services in clouddev.yml.

6. Web Dashboard

Open http://localhost:4580 to see live service status:

![CloudDev Dashboard showing all services running in green]

How it compares to LocalStack

Feature	CloudDev	LocalStack Free
Requires account	❌ No	✅ Yes (now required)
Open source	✅ Yes	⚠️ Partial
S3, DynamoDB, Lambda	✅ Yes	✅ Yes
Web dashboard	✅ Yes	❌ No
IaC auto-detection	✅ Yes	❌ No
Cost	🆓 Free forever	💰 Paid plans

Tech stack

CloudDev is built in Go — which means:

Single binary, no runtime dependencies
Fast startup
Low memory usage

What's next

Kubernetes integration
Cost estimation before deployment
Plugin system for additional AWS services
More AWS service emulators

Try it out

👉 GitHub: github.com/Jeffrin-dev/CloudDev

If you find it useful, please ⭐ the repo and share it with
other AWS developers who are affected by the LocalStack change!

Contributions are very welcome — check out
CONTRIBUTING.md
to get started.

Built with Go, Let's Get Started !

Claude AI Gets Weirdly Slow After 9 PM (I Noticed It While Reviewing Code)

Twisted-Code'r — Fri, 13 Mar 2026 15:01:15 +0000

I ran into something interesting recently while using Claude AI to review some of my code.

During the day the responses were fast. I could paste a file, ask for suggestions, iterate quickly, and the workflow felt smooth.

But when I tried doing the same thing later in the evening — around 9 PM and after — the experience changed a lot.

Responses suddenly took much longer.

Sometimes it would sit there “thinking” for quite a while before returning the review.

At first I assumed it was something on my side:

maybe my internet
maybe the browser tab
maybe I pasted too much code

But after trying a few times on different days, I noticed the same pattern.

It’s Probably Peak Usage

My guess is pretty simple.

Evenings are when a lot of developers start working on:

side projects
open source
debugging issues from the day
experimenting with AI tools

So tools like Claude, ChatGPT, and Gemini probably get a huge spike in requests around that time.

And when millions of prompts hit the system at once, response times naturally slow down.

Why Developers Notice It More

The funny part is that developers probably notice this more than anyone else.

When you're using AI for code review or debugging, you're usually sending multiple prompts in a row:

ask for review
ask for improvements
clarify something
test another idea

That workflow depends on fast feedback.

Even a small delay starts to feel frustrating when you're in that loop.

Not Really a Complaint

To be fair, these systems are doing a massive amount of work behind the scenes. Running models at this scale is insanely expensive.

But it was interesting to notice how time of day actually affects the experience.

Now I’m curious if other developers have noticed the same thing when using AI tools for code reviews at night.

Is it just me, or does AI also have a “rush hour”?

I had a problem that I suspect many developers have — duplicate files scattered across external drives, backup archives, and media collections with no reliable way to clean them up without risking data loss. Found a way, moreover made a way..

Twisted-Code'r — Fri, 13 Mar 2026 14:39:48 +0000

How I Built a Duplicate File Finder That Handles 1 Million Files Without Crashing

Twisted-Code'r ・ Mar 11

#python #opensource #cli #tutorial

DEV Community: Twisted-Code'r

AUTO WRITES PR FOR YOU - PRDRAFT

I got tired of PRs that just said "fix" — so I built a GitHub App that writes the description for you

I got tired of PRs that just said "fix" — so I built a GitHub App that writes the description for you

The real problem isn't laziness

What I built

Stack (all free)

The thing nobody tells you about building GitHub Apps

Where it is now

Try it

Review

I built a GitHub App that auto-writes PR descriptions — here's what happened in 8 days

Day 1–2: Does anyone actually care?

Day 3–5: Building the thing

Day 5: First external user

Day 6–7: Bug fixes and distribution

Day 7: Dashboard

Day 8: More dashboard improvements

The payment provider saga

Current numbers (Day 8)

What's next

Try it

I stopped writing PR descriptions. Here's what I did instead.

I built an open-source tool that stops personal data from leaking into AI chatbots

CloudDev v0.3.0 — I built a free LocalStack alternative in Go with 17 AWS services published: true

Why I Built This

What's in v0.3.0

🔐 Auth & Security

🏗️ Orchestration

🌉 Events & Cache

Quick Start

Works With Your Existing AWS CLI

The Web Dashboard

How It Was Built

Complete Service & Port Reference

What's Next

Get Involved

I Built a Free Voice-Enabled AI Agent That Uses Notion as Its Brain — Notion MCP Challenge

What I Built

The Problem I Was Solving

The Free Stack (Genuinely $0/Month)

How It Works — The MCP Architecture

Feature Walkthrough

1. Smart Save

2. Natural Language Routing

3. Semantic Search

4. Think — Multi-hop Reasoning

5. Recall — Knowledge Evolution

6. The Dashboard

7. AI Insights

8. Knowledge Graph

9. Suggest — Auto Related Notes

10. Full Telegram Bot

11. Todo List

12. Voice Mode

The Hardest Parts to Build

Semantic search that doesn't hallucinate

The suggest feature

Command routing strictness

What I'd Add Next

Links

Try It Yourself

I Ditched GitHub Copilot and Now I Code With AI for Free — Locally

What You Actually Need

Step 1: Install Ollama

Step 2: Pull a Coding Model

For most people (8GB RAM):

If you have 16GB+ RAM:

If you're on Apple Silicon (M1/M2/M3):

Absolute minimum (4GB RAM):

Step 3: Pick Your Editor Integration

VS Code → Continue

Neovim → gen.nvim or avante.nvim

JetBrains → Continue (same plugin, different install)

Step 4: Supercharge It With Open WebUI (Optional but Worth It)

Real-World Performance

The Part Nobody Tells You

Bonus: Free Cloud Options When Local Isn't Enough

TL;DR

CloudDev v0.2.0 — SNS, Secrets Manager, CloudWatch Logs and more!