LangSmart Smartflow · Interactive Guide

SafeChat v3 Usage Guide

Step-by-step walkthrough of every new capability — from model selection to agent skills, team channels, and beyond.

Claude 4.6 · Opus & Sonnet Adaptive Thinking Code Execution Prompt Library Agent Skills Team Channels
Step 01 · Getting Started
Choose Your Model
The model selector at the top of the chat gives you access to the full provider lineup — including the brand-new Claude 4.6 Opus and Sonnet, which support adaptive thinking. The UI automatically adjusts available features based on what each model supports.
chat.langsmart.app
MODEL
★ Claude Opus 4.6
Claude Sonnet 4.6
Claude Sonnet 4.5
GPT-5.2
Gemini 2.5 Pro
Anthropic — Claude 4.6
Opus 4.6 NEW
Adaptive thinking · 1M ctx
Sonnet 4.6 NEW
Adaptive thinking · coding
OpenAI — GPT-5
GPT-5.2 Flagship
Images · PDF text · Docs
GPT-5 Mini
Fast · cost-efficient
Google — Gemini
Gemini 2.5 Pro
Video · Audio · PDF native
Gemini 2.5 Flash
Fast multimodal
💡 Auto-capability detection
Switching models automatically shows or hides the Thinking button, Tools toolbar, and adjusts which file types are accepted.
⭐ Claude 4.6 is new
Opus 4.6 and Sonnet 4.6 were released Feb 2026 and use adaptive thinking — a smarter mode than the older extended thinking.
2
Step 02 · Advanced Reasoning
Enable Thinking Mode
When a Claude model is selected, a Think button appears in the toolbar. Click it to activate reasoning — Claude will show its internal thought process before answering. For Claude 4.6 models this is adaptive (Claude decides depth). For 4.5 it's extended (you set a token budget).
chat.langsmart.app
Claude Opus 4.6
Think · Deep ▾
Adaptive mode active
Design a fault-tolerant microservices architecture for a payment processing system with 99.99% uptime.
🧠 Thinking
Let me think through the key concerns: payment systems need ACID guarantees, idempotency for retry safety, circuit breakers to isolate failures, event sourcing for audit trails, and distributed saga patterns for multi-service transactions...
Based on my analysis, here's a fault-tolerant architecture: 1) Saga pattern for distributed transactions, 2) Event sourcing + CQRS for audit, 3) Circuit breakers on every service boundary...
Ask a follow-up...
🔵 Adaptive vs Extended
Adaptive (Claude 4.6): Claude decides how much to think. Extended (Claude 4.5): you pick Deep / Balanced / Fast effort level.
👁 Thinking blocks
The purple reasoning block shows Claude's internal scratchpad — reasoning through trade-offs before committing to an answer. Toggle to collapse.
3
Step 03 · Tools Toolbar
Code Execution & Web Search
The tools toolbar sits above the chat input. Toggle Code to let Claude write and run Python in a sandboxed container, or Search to give Claude live web access. Enabled tools are highlighted in green — disabled ones are greyed out.
chat.langsmart.app
Claude Sonnet 4.6
Analyse this sales dataset and plot the top 5 products by revenue.
⚙ Code Execution
# Load and analyse sales data
import pandas as pd
import matplotlib.pyplot as plt
df = pd.read_csv('sales.csv')
top5 = df.groupby('product')['revenue'].sum().nlargest(5)
✅ Executed · 0.4s · Chart saved to output
Tools
⚙ Code
🔍 Search
Gemini: images · PDF · audio · video · docs
Type a message...
⚙ Code tool
Claude writes Python, runs it in a sandbox, and returns results inline. Great for data analysis, file processing, math, and charting.
🔍 Search tool
Gives Claude real-time web access. Answers about current events, pricing, docs, or anything that changes over time.
⚠ Model support
Code execution requires Claude 4.x or 4.5+ models. The toolbar auto-hides tools that the selected model doesn't support.
4
Step 04 · Prompt Library
Browse, Build & Save Prompts
Click Prompt Library in the left sidebar to open the modal. Browse curated and community prompts, use the Workbench-style Builder to craft new ones with AI scoring, then submit to share with your team. Prompts are stored in MongoDB and persisted across sessions.
chat.langsmart.app — Prompt Library
💡 Use in Chat
Clicking "Use in Chat →" sets the prompt as your system prompt — a purple indicator appears in the chat header so you always know it's active.
🏗 Builder tab
Enter your intent, let AI generate a system + user prompt pair, preview it live, then score it with Claude before saving to the library.
☁️ Persisted in MongoDB
All submitted prompts are stored server-side. No localStorage — your team shares the same library across devices and sessions.
5
Step 05 · Agent Skills
Activate Specialised AI Personas
The Agent Skills tab (inside the Prompt Library modal, also in the sidebar) lets you apply structured behavioural instructions to the AI. Pick from seeded Smartflow skills or write your own using the workbench editor. Skills are stored in MongoDB and can be activated instantly in any chat.
chat.langsmart.app — Agent Skills
Browse Skills
+ Create Skill
🔑 What is a skill?
A structured SKILL.md file that tells the AI how to behave — its persona, tools, reasoning approach, and output format — for a specific domain.
✍️ Create your own
The "Create Skill" workbench has a full code editor with live preview. Load a template, customise it, then save to MongoDB for your whole team.
⚡ Instant activation
Clicking Activate sets the skill as the active system prompt immediately. A purple indicator in the chat header confirms it's live.
6
Step 06 · Team Collaboration
Team Channels & Threads
Click any channel in the sidebar to see all its threads. Each thread card now shows participant avatars — mini coloured circles with initials for everyone who has posted. Click a thread to jump straight into that conversation.
chat.langsmart.app — #engineering
← Back
# engineering
4 threads
+ New Thread
API Gateway Migration Plan
3 new
Should we migrate the legacy API gateway to Kong or AWS API Gateway? Here are my initial thoughts...
Redis Caching Strategy
We need to decide on TTL values for the semantic cache layer and whether to enable key expiry...
Q2 Performance Benchmarks
Attaching the Grafana export. P99 latency is up 15% — let's trace the regression to the new proxy layer...
👥 Participant avatars
Each thread card shows coloured avatar circles for everyone who has posted — hover to see the full name. Shows up to 6 participants.
🔔 Unread badges
Blue "N new" badges on threads show unread messages since your last visit, so you never miss an important discussion.
7
Step 07 · Multimodal Attachments
File Uploads & Smart Warnings
Attach images, PDFs, audio, video, and documents using the paperclip button. The chat knows exactly what each model supports — if you attach an unsupported file type, a warning toast appears immediately telling you what to do, instead of silently failing.
chat.langsmart.app
Provider / Model Images PDF Audio Video Docs / Code
Claude 4.6 / 4.5 ✓ Native ✓ Native ✗ No ✗ No
GPT-5 / GPT-4o ✓ Native ⚠ Text extract ✗ No ✗ No
GPT-4o Audio ✓ Native
Gemini 2.5 Pro/Flash ✓ Native ✓ Native ✓ Native ✓ Native
⚠️
Video files are not supported by Claude Opus 4.6. Switch to Gemini 2.5 Pro for native video analysis, or the file will be noted as unsupported.
🌟 Use Gemini for video/audio
Neither Claude nor GPT support video input in the chat API. Gemini 2.5 Pro/Flash is the only provider with native video and audio analysis.
📄 PDF note
Claude reads PDFs natively (full layout + structure). GPT extracts text only. For best PDF results, use a Claude model.
🔔 Non-blocking toasts
Warnings appear as animated toasts at the bottom of the screen and auto-dismiss after 7 seconds. You can still send the message.
Quick Reference
Everything at a Glance
1️⃣ Pick Claude 4.6
New Opus/Sonnet 4.6 in the dropdown for best adaptive reasoning and coding performance.
2️⃣ Hit Think
Toggle the Think button for complex questions. Watch Claude reason before answering.
3️⃣ Enable Code tool
Turn on Code + Search pills in the toolbar to unlock live execution and web search.
4️⃣ Load a Prompt
Open Prompt Library → Browse → click "Use in Chat" for an instant expert system prompt.
5️⃣ Activate a Skill
Skills tab → pick Financial Analyst, Code Review Pro, etc. for deep-domain AI expertise.
6️⃣ Collaborate in channels
Use team channels for shared threads. See who's participated via avatar stacks at a glance.