The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
I've used LLMs for 3 years but never truly understood them until a recent flight without internet.
I watched Andrej Karpathy's breakdown. As an OpenAI co-founder, he makes it click.
The big realization: LLMs are text predictors.
Long contexts (huge chat history) make accurate prediction harder. This is a major reason why hallucinations happen.
I used to cram everything into one giant thread. I didn't realize I was breaking the model.
My new approach:
Separate chats for different tasks.
Short, concrete instructions.
Since most models use similar training data, the real difference is your prompting.
Short context = Better outputs.
If you want to learn more, here is the video:
https://www.youtube.com/watch?v=zjkBMFhNj_g
Exactly. Keeping context short has been a game changer.
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Challenges
View all
Related posts
AI Agent Integration & Custom Website Chat Widget
AI Lead Qualification & Appointment Automation System
AI Lead Conversion Platform — Telegram, CRM & Automated Scheduling
Designed and developed an AI-powered lead conversion system that handles incoming Telegram conversations from first contact to qualification, CRM synchronization, and appointment booking.
The AI works with conversation history, lead profile data, unanswered questions, current funnel stage, available actions, and verified CRM data to determine the next appropriate step and generate structured responses.
The system automatically creates and updates leads in the CRM, manages qualification stages, books meetings, generates Yandex Telemost links, sends reminders, handles rescheduling, and escalates exceptional cases to a human manager.
The architecture includes dialogue state management, message deduplication and batching, PostgreSQL persistence, concurrency protection, validation before actions are executed, and safeguards against sending unverified or prohibited information.
This is not just an AI chatbot — it is an automated sales workflow that connects Telegram, AI, CRM, scheduling, and human escalation into one controlled process.
AI makes decisions within controlled business rules instead of operating as an unrestricted chatbot.
The concurrency safeguards and validation-before-actions detail is a strong distinction here—booking systems usually fail at the edges, not in the happy path. I’d expose a compact lead state to the operator (stage, next action, last unanswered question) so a human can resume...
I just open-sourced my Contra MCP Starter Toolkit ⚡
Contra now has a native Model Context Protocol server, which means MCP-compatible AI agents can work directly with your Contra account — profile, services, portfolio, proposals, projects, invoices and more.
The problem: getting a coding agent properly connected still involves OAuth, PKCE, token refresh, MCP configuration and making sure the agent respects Contra’s prepare → confirm workflow.
So I packaged the setup I was using into an open-source starter kit:
What’s inside:
• One-command OAuth setup — run npm run auth, authorize Contra in the browser, and the local callback handles PKCE and token exchange.
• Zero dependencies — native Node.js 18+ only: fetch, crypto, http and fs.
• Automatic token refresh — the client catches expired sessions, refreshes the access token and retries the MCP request.
• Safe agent workflows — instructions for Cursor, Google Antigravity, Claude Code and Windsurf built around Contra’s native prepare → confirm protocol, including a full before/after review before any live mutation.
• Zero-code AI mode — copy-paste prompts for inspecting your Contra account, restructuring services, updating your profile and publishing portfolio case studies without writing JavaScript yourself.
• Content-as-Code templates — Markdown blueprints for profile positioning, flagship services, pricing, FAQs and structured case studies.
The interesting part for me isn’t just Contra.
It’s the workflow: your professional profile becomes something an AI coding agent can inspect, reason about and maintain through structured tools — while you keep final approval over every write.
If you use Contra + Cursor / Claude Code / Antigravity / Windsurf, try it and tell me what’s missing.
