Practical techniques, workflows, and tool guides — everything you need to get more from AI today.
18 Guides
4 Categories
Always Free
All
🧠 Prompting
⚙️ Workflows
🤖 Models
🛠 Tools
🧠 PromptingBeginner
Chain-of-Thought Prompting
Force AI to reason step-by-step instead of jumping to an answer. Dramatically improves accuracy on maths, logic, and multi-step problems.
🕐 5 min read
Key Technique
Add "Think step by step" or "Let's reason through this carefully" to any complex prompt. Works on all major models.
Chain-of-thought (CoT) prompting was formally studied by Google researchers in 2022 and remains one of the most reliable techniques. Simply asking the model to "think step by step" before answering reduces errors on arithmetic and logic tasks by 30–50% on benchmark tests.
Example prompt:
"A store sells 3 types of apples at £1.20, £1.50, and £2.00 each. I buy 4 of the first, 3 of the second, and 2 of the third. What's my total? Think step by step."
Why it works: LLMs generate tokens sequentially — by making the model articulate intermediate steps, it's less likely to "jump" to a wrong conclusion. The reasoning tokens act as working memory.
Pro tip: For very complex problems, ask "Let's think through this carefully, considering all factors" — this triggers even deeper reasoning than the basic phrase.
🧠 PromptingBeginner
Few-Shot Prompting
Show AI 2–3 examples of exactly what you want, then ask for your actual output. Consistently outperforms zero-shot prompts for structured tasks.
🕐 4 min read
Key Technique
Before your real request, add "Here are some examples:" followed by 2–3 input/output pairs in the exact format you want.
Few-shot prompting is one of the highest-ROI techniques — it takes 60 seconds to write but often doubles output quality.
Example:
Input: "Write a subject line for a cold email about project management software"
Output 1: "Cut your team's meeting time in half"
Output 2: "The project tracker your team will actually use"
Output 3: "How [Company] reduced missed deadlines by 60%"
Now write a subject line for: [your product]
Key insight: The examples don't need to be perfect — they just need to show the model the format, tone, and length you want. Even 1–2 examples beat zero-shot for most structured tasks.
When to use it: Subject lines, product descriptions, social posts, data extraction, classification tasks.
🧠 PromptingIntermediate
The RISEN Framework
A 5-part prompt structure used by professional prompt engineers. Stands for Role, Instructions, Steps, End Goal, Narrowing.
🕐 6 min read
Key Technique
Structure every serious prompt as: Role → Instructions → Steps → End Goal → Narrowing (constraints). This single framework covers 90% of professional use cases.
RISEN is the most practical all-purpose prompting framework. Here's how each part works:
R — Role: "You are a senior financial analyst with 15 years of experience in SaaS valuations."
I — Instructions: "Analyse this company's revenue metrics and write an investment memo."
S — Steps: "First, assess growth rate trends. Then evaluate net revenue retention. Finally, benchmark against public SaaS multiples."
E — End Goal: "The output should be a structured 400-word memo suitable for a partner meeting."
N — Narrowing: "Do not speculate about future fundraising. Focus only on historical data provided."
Full example prompt:
"You are a senior financial analyst. Analyse this company data [paste data] and write a 400-word investment memo. First assess growth, then retention, then comparable multiples. The memo is for partner review. Do not speculate beyond the data."
Why it works: Each element eliminates a different type of model error — role sets expertise, instructions set task, steps set process, end goal sets format, narrowing prevents hallucination.
🧠 PromptingIntermediate
Structured JSON Output
Get AI to return clean, parseable JSON every time — essential for building AI-powered apps, automations, and data pipelines.
🕐 5 min read
Key Technique
End your prompt with "Return ONLY valid JSON. No markdown, no explanation. Schema: { field1: type, field2: type }". Current Claude, ChatGPT and Gemini models are near-100% reliable with this, and their APIs also offer structured-output modes that guarantee valid JSON.
Getting reliable structured output is the difference between a prototype and a production system.
Template that works:
"Extract the following from this job posting: [job text]
Return ONLY valid JSON with this exact schema:
{
"title": "string",
"company": "string",
"location": "string",
"salary_min": number or null,
"salary_max": number or null,
"remote": boolean,
"requirements": ["string"]
}
No markdown. No explanation. Just the JSON."
Model comparison: current frontier models from Anthropic, OpenAI and Google handle this very reliably, and API structured-output modes remove the remaining failures. Small local models need more hand-holding, so validate their output.
For Anthropic's API specifically: use the response_format parameter or the new tool use / structured outputs feature for guaranteed-valid JSON at the API level.
🧠 PromptingAdvanced
Prompt Chaining
Break complex tasks into sequential prompts where each output feeds the next. Produces dramatically better results than one giant prompt.
🕐 8 min read
Key Technique
Map your task as a pipeline: generate → critique → refine → format. Each stage is a separate API call with the previous output as context.
Prompt chaining is how professional AI applications produce consistently excellent output. Instead of one massive prompt, you build a pipeline:
Example — Writing a blog post:
Step 1 (Research): "List 10 key facts about [topic] from first principles. Be precise and cite specific data where possible."
Step 2 (Outline): "Using these facts [paste step 1 output], create a detailed outline for a 1,200-word blog post for a technical audience."
Step 3 (Draft): "Write section 2 of this outline [paste section] in full. Target 300 words, active voice, no filler."
Step 4 (Critique): "Review this draft for clarity, accuracy, and engagement. List specific improvements."
Step 5 (Refine): "Rewrite with these improvements applied: [paste critique]"
Why not one big prompt? LLMs degrade in quality over long outputs. A focused 200-word output is almost always better than asking for 1,000 words in one go.
Tools that make this easier: LangChain, Anthropic's API batching, OpenAI Assistants API with threads.
⚙️ WorkflowsBeginner
Write 10x Faster with AI
A practical AI writing workflow used by professional content creators. From blank page to polished draft in under 20 minutes.
🕐 5 min read
Key Technique
Never ask AI to "write an article". Instead: brainstorm → outline → section-by-section drafting → edit in your voice.
The mistake most people make: asking ChatGPT to "write a 1,000-word article about X" and getting generic slop.
The professional workflow:
1. Brainstorm (2 min): "Give me 10 angles for an article about [topic] targeting [audience]. What would surprise them or challenge assumptions?"
2. Pick your angle and outline (3 min): "Create a detailed outline for this angle: [chosen angle]. 5 sections, each with a specific insight or example."
3. Draft section by section (10 min): "Write section 1 [paste section outline]. 250 words. Use active voice. Include a specific example. Do not use filler phrases like 'In conclusion' or 'It is worth noting'."
4. Edit in your voice (5 min): Paste your draft back: "Edit this to sound more like [describe your voice — direct, conversational, slightly sarcastic]. Keep all the facts. Tighten every sentence."
Result: A draft that sounds like you, not like a robot. The key is that you're directing and editing — AI is handling the mechanical generation work.
⚙️ WorkflowsBeginner
Build a Personal Research Pipeline
Use Perplexity + Claude (or ChatGPT) to research any topic thoroughly in 15 minutes instead of 3 hours.
🕐 5 min read
Key Technique
Use Perplexity for current facts and sources, then paste those findings into Claude for synthesis, analysis, and writing.
Most people use AI for research wrong — they ask ChatGPT questions and get hallucinated answers. Here's a reliable pipeline:
Step 1 — Gather facts with Perplexity: Perplexity is grounded in the web with citations. Ask: "What are the latest developments in [topic]? Give me key stats, recent events, and important players. Include sources." Save the response.
Step 2 — Synthesise with Claude: Paste Perplexity's output into Claude and say: "Based on these research notes, what are the 3 most important insights? What's counterintuitive? What are the implications for [your use case]?"
Step 3 — Go deep on what matters: Take the most interesting thread from step 2 and ask Claude: "Explain [specific concept] in depth. Use analogies. Cover edge cases and common misconceptions."
Step 4 — Sanity check: "What important perspectives or counterarguments am I missing? What would a critic of this view say?"
Why this works: Perplexity handles recency and citation, Claude handles reasoning and synthesis. Each model does what it's best at.
⚙️ WorkflowsIntermediate
Code Review with AI
A structured workflow for using AI to catch bugs, security issues, and performance problems before code goes to production.
🕐 6 min read
Key Technique
Don't ask AI to "review my code". Give it a specific checklist: security, performance, edge cases, readability — one at a time.
Vague code review requests get vague responses. Here's a structured approach:
Security review prompt:
"Review this code for security vulnerabilities. Focus specifically on: SQL injection, XSS, authentication bypass, insecure deserialization, and exposed secrets. For each issue found, show the vulnerable line and a fixed version."
[paste code]
Performance review prompt:
"Analyse this code for performance issues. Look for: N+1 queries, unnecessary re-renders, missing indexes, synchronous operations that should be async, and memory leaks. Show specific line numbers."
Edge case review:
"What inputs or states would cause this code to fail or behave unexpectedly? Consider: null/undefined inputs, empty arrays, very large datasets, concurrent requests, and network failures."
Best models for code review: any current frontier model works well. Claude is especially strong at reasoning about code logic, ChatGPT at spotting security issues, and Gemini at very large codebases thanks to its long context window. Try two on the same diff; they catch different things.
Pro tip: Cursor IDE does this automatically as you write, which makes its paid plan worthwhile for serious developers.
⚙️ WorkflowsBeginner
Automate Your Morning Briefing
Set up a daily AI workflow that delivers a personalised summary of what matters to you — in under 10 minutes to configure.
🕐 4 min read
Key Technique
Use a custom GPT or Claude Project with your news sources pre-loaded. Ask it every morning: "What happened overnight in [your field] that I need to know?"
A practical morning routine that takes 3 minutes:
Option A — Claude Projects (simplest):
1. Create a new Claude Project
2. In the Project instructions, write: "You are my daily AI briefing assistant. When I say 'morning briefing', summarise the key AI developments from the past 24 hours, what they mean for practitioners, and 1 thing I should do or learn today. Be concise — 5 bullet points max."
3. Each morning: open the project, type "morning briefing"
4. Claude uses its knowledge cutoff — for truly real-time, paste headlines from npulse.io
Option B — Custom GPT with web browsing:
1. Create a Custom GPT with web browsing enabled
2. System prompt: same as above
3. This version can actually browse current news
Option C — Perplexity Daily Digest:
Perplexity has a built-in "Daily Digest" feature (paid plans) that emails you a personalised briefing.
For AI news specifically: Bookmark npulse.io and spend 5 minutes scanning headlines. Use AI to go deeper on anything that catches your eye.
⚙️ WorkflowsIntermediate
Use AI as a Thinking Partner
The most underused AI skill — using it to pressure-test your ideas, spot blind spots, and make better decisions.
🕐 6 min read
Key Technique
The "Devil's Advocate" prompt. Share your plan, then ask AI to argue against it as strongly as possible. The objections it raises are usually the ones you haven't considered.
Most people use AI to validate ideas. The highest-value use is the opposite — challenging them.
Devil's Advocate prompt:
"Here is my plan: [describe plan]. I want you to argue against this as strongly as possible. What assumptions am I making that could be wrong? What could go wrong that I haven't considered? What would a smart critic say? Be specific, not generic."
Pre-mortem prompt:
"Imagine it's 12 months from now and this project has completely failed. Walk me through the most plausible reasons why. What were the early warning signs we ignored?"
Decision audit:
"I'm deciding between [option A] and [option B]. Here are my current reasons for leaning toward A: [list]. What are the strongest arguments for B that I might be discounting? What information should I gather before deciding?"
Why this is powerful: AI has no ego investment in your idea. It won't soften criticism to spare your feelings. A well-prompted critique session takes 10 minutes and regularly surfaces issues that would have taken months to discover in practice.
🤖 ModelsBeginner
Claude vs ChatGPT vs Gemini: Which to Use When
A practical breakdown of when each model excels — not benchmark scores, but real-world task performance.
🕐 7 min read
Key Technique
Use Claude for long documents and careful writing, ChatGPT for broad all-round tasks and images, and Gemini when you live in Google’s apps or need to process very large files.
Stop asking which model is "best". Ask which is best for your task.
Claude (Anthropic) — Best for:
- Long-form writing that sounds human
- Complex reasoning and analysis
- Following nuanced instructions precisely
- Coding with careful explanation
- Sensitive topics (most thoughtful responses)
Weak at: Very recent events (knowledge cutoff), image generation
ChatGPT (OpenAI) — Best for:
- General-purpose tasks (Jack of all trades)
- Code generation and debugging
- Built-in image generation, including accurate text on posters
- Agents, apps and integrations
- Custom GPTs
Weak at: Sometimes verbose, can be overconfident
Gemini (Google) — Best for:
- Processing very large documents (long context window)
- Multimodal tasks (image, video, audio)
- Google Workspace integration
- Free tier is genuinely good
Weak at: Can feel less "sharp" on pure text reasoning vs Claude
Quick decision guide:
- Writing an email, report, or article → Claude
- Building a quick app or script → ChatGPT or Claude
- Analysing a 500-page PDF → Gemini or NotebookLM
- Research with web access → Perplexity (not a pure LLM but often the right answer)
- Posters and images with text → ChatGPT or Gemini (both handle Hindi text well now; see our test)
- Artistic images → Midjourney
- Local and free → Stable Diffusion or Flux
Updated October 2026.
🤖 ModelsBeginner
Best Free AI Tools Right Now
The best AI tools you can use today without spending a penny — and what each one is actually good for.
🕐 5 min read
Key Technique
Claude.ai free tier, Gemini free tier, and Perplexity free tier cover 80% of what most people pay for. Use the right free tool for each job.
You don't need a paid subscription to get serious value from AI. Here's the best free stack:
Writing & reasoning: Claude.ai (free tier). Strong at long documents and careful writing; daily message limits apply.
Research with sources: Perplexity.ai (free) — web-grounded answers with citations. The free tier is excellent for research.
General + Google integration: Gemini (free) — strong free tier, integrates with Google Docs, Gmail, and Drive.
Images: ChatGPT and Gemini (both free with daily limits, and both get text right, including Hindi; see our test), and Canva's free plan for layouts
Code: GitHub Copilot (free tier), Cursor (free tier), Gemini Code Assist (free for individuals)
Documents/PDF chat: NotebookLM (free; answers only from your files, with citations), Claude (upload PDFs on the free tier)
Audio transcription: Whisper via Groq (near-instant, free) or Otter.ai (free tier)
Video: Runway (free credits), Kling (free tier)
Pro tip: Before paying for any AI subscription, spend one week using the free tiers of Claude, Gemini, and Perplexity. Most people find they only need to pay for 1–2 tools once they know which fits their workflow.
🤖 ModelsIntermediate
Local AI vs Cloud AI
When to run AI models on your own hardware — privacy, cost, speed, and what's actually worth running locally.
🕐 6 min read
Key Technique
Use local models (via Ollama) for anything involving private data. Use cloud models for tasks requiring the latest capabilities. The gap has narrowed dramatically.
Running AI locally has gone from niche to practical. Here's the honest breakdown:
Run locally when:
- Processing sensitive data (medical records, legal docs, private business data)
- Cost matters at scale (local = near-zero marginal cost after hardware)
- You need offline capability
- Latency is critical (local can be faster for short queries)
Use cloud when:
- You need frontier capabilities (local models are still 6–12 months behind on complex reasoning)
- You're on a laptop without a decent GPU
- Task volume is low (cloud is cheaper than buying hardware for occasional use)
Best local model families right now (new versions arrive every few weeks, so check the Ollama library for the latest):
- Qwen (Alibaba): the 8–9B versions are excellent all-rounders and fit a 16GB laptop
- Gemma (Google): small versions run on any laptop; mid-size versions also understand images
- Llama (Meta): well documented; the largest versions need 40GB+ of memory
- Phi (Microsoft): tiny models with an excellent quality-to-size ratio
How to get started: Install Ollama (ollama.com, free and open source). Run `ollama run qwen3:8b` in a terminal. The first run downloads the model (about 5GB); after that it starts in seconds and works offline.
Hardware reality: 16GB of RAM comfortably runs 7–9B models; a Mac with 32GB+ unified memory or an NVIDIA GPU with 16–24GB of VRAM handles much larger ones.
Updated October 2026.
🛠 ToolsBeginner
Cursor AI: Code Faster Than You Ever Have
The AI coding assistant that's replaced VS Code for hundreds of thousands of developers. Here's the fastest way to get productive.
🕐 5 min read
Key Technique
Use Cursor's Composer (Cmd+I) for multi-file changes, not just autocomplete. Describe what you want to build and watch it edit multiple files at once.
Cursor is VS Code with AI deeply integrated — it reads your entire codebase, not just the file you're in.
The 3 features that matter most:
1. Composer (Cmd+I) — Multi-file AI editing. Describe a change in plain English: "Add user authentication to this Express app using JWT. Update the routes, add middleware, and create a User model." Cursor edits all relevant files simultaneously and shows a diff.
2. Chat (Cmd+L) — Ask questions about your codebase: "Why is this function slow?" or "What does this regex do?" It reads your full project context.
3. Inline editing (Cmd+K) — Select code, press Cmd+K, describe a change: "Refactor this to use async/await instead of callbacks". Instantly rewrites the selection.
Setup tips:
- Add project rules (in .cursor/rules; older projects use a .cursorrules file) with your coding standards and preferred patterns. Cursor reads these as context for every request
- Use the model picker to switch between models from Anthropic, OpenAI and Google: a strong reasoning model for hard changes, a fast one for small edits
Pricing: the free tier has limited usage. The paid plan is worth it for daily use — it becomes like having a senior developer pair-programming with you.
Best for: Web developers, Python/data science, anyone who writes code daily.
🛠 ToolsBeginner
Perplexity AI: Research Without the Rabbit Holes
Perplexity answers questions with real sources, cited inline — no hallucination guessing game. Here's how to get the most out of it.
🕐 4 min read
Key Technique
Use Follow-up questions in the same thread — Perplexity remembers context. Ask "go deeper on X" or "what does source 3 say about Y" to drill down without losing thread.
Perplexity is what Google Search should be — an answer engine that shows its work.
When to use Perplexity instead of ChatGPT:
- Any question about current events or recent data
- Anything where you need to verify sources
- Research that you'll act on professionally
- Fact-checking claims
Power user tips:
1. Focus mode: Switch from "All" to "Academic" for research papers, "YouTube" for video results, or "Reddit" for community opinions. Massively improves result quality for specific needs.
2. Pro Search: Asks clarifying questions before searching, resulting in much more targeted answers. Toggle it on in settings (free users get 5/day, Pro gets unlimited).
3. Collections: Save related searches to a collection and Perplexity builds a knowledge base you can query.
4. Use it for live product research: "Compare the top 3 project management tools for remote teams under $20/user/month. What do users say are the main complaints about each?"
Free vs Pro: The free tier is excellent for most users. Pro adds unlimited Pro Search, file uploads, and access to frontier models from OpenAI, Anthropic and Google.
🛠 ToolsBeginner
NotebookLM: Turn Any Document into an Expert
Google's NotebookLM lets you chat with your own documents — PDFs, notes, URLs — and get cited, accurate answers. It only answers from what you give it.
🕐 4 min read
Key Technique
Upload 5–10 sources on the same topic, then ask for a synthesis across all of them. NotebookLM will find connections, contradictions, and themes you'd spend hours finding manually.
NotebookLM is unique: it only uses the sources you provide, so there's zero hallucination risk outside your documents.
Best use cases:
1. Research synthesis: Upload 10 papers, ask: "What do all these sources agree on? Where do they contradict each other? What questions remain unanswered?"
2. Due diligence: Upload a company's annual report, investor deck, and press releases. Ask: "What are the key risks mentioned? How does the investor deck's narrative differ from the annual report?"
3. Study/learning: Upload your textbook PDFs or lecture notes. Ask it to quiz you, explain concepts differently, or generate flashcards.
4. Legal/contract review: Upload contracts (make sure you're not violating confidentiality). Ask: "What are the key obligations on my side? What are the termination conditions? What's unusual compared to standard terms?"
The Audio Overview feature: NotebookLM can generate a podcast-style conversation between two AI hosts discussing your documents. Genuinely useful for absorbing long material while commuting.
Limitation: It's only as good as what you give it. Garbage in, garbage out — make sure your source documents are high quality.
🛠 ToolsBeginner
ElevenLabs: Clone Your Voice in 60 Seconds
Create a perfect digital clone of your voice with under 2 minutes of audio. Use it for content creation, voiceovers, and accessible text-to-speech.
🕐 3 min read
Key Technique
Record your voice sample in a quiet room, speaking naturally with varied intonation — not reading flatly. Quality of the sample directly determines quality of the clone.
ElevenLabs has the most realistic AI voice generation available. Here's how to set up a voice clone:
Voice cloning (Instant Clone — free tier):
1. Go to ElevenLabs.io → My Voices → Add Voice
2. Record or upload 1–2 minutes of clear audio (your voice, minimal background noise)
3. Name your voice and submit
4. Your clone is ready in ~30 seconds
Tips for better clones:
- Record in a quiet room (not echoey)
- Vary your intonation naturally — don't read monotonously
- Include some emotional range: explain something enthusiastically, ask a question, make a point firmly
- 3–5 minutes of audio gives noticeably better results than 1 minute
Best uses:
- Podcast/YouTube content without re-recording
- Creating audio versions of your blog posts
- Consistent voice across multiple videos
- Accessibility features for your product
Professional Clone: ElevenLabs' Professional Voice Clone (paid) uses 30+ minutes of audio and produces broadcast-quality results indistinguishable from a real recording.
Pricing: there's a free tier to try it out; paid plans add more characters per month and higher-quality voice cloning.
🛠 ToolsIntermediate
Midjourney: Create Stunning Images with the Right Prompts
Midjourney v6 produces photorealistic images and stunning artwork — but 90% of the quality comes from how you write the prompt.
Midjourney is the current gold standard for AI image generation. Here's the prompt structure used by top creators:
The formula: [Subject + action] [artistic style] [lighting] [composition] [technical params]
Example (bad): "a person in a city"
Example (good): "Young professional walking through Tokyo at golden hour, street photography style, shallow depth of field, shot on 35mm, cinematic lighting, --ar 16:9 --style raw --v 6"
Key parameters:
- --ar 16:9 or --ar 1:1 — aspect ratio
- --v 6 — use the latest model version
- --style raw — removes Midjourney's auto-beautification for more realistic output
- --q 2 — higher quality (slower)
- --no [thing] — exclude elements (e.g., --no text, watermark)
Style modifiers that work consistently:
- "shot on Sony A7IV, 85mm lens" → adds photographic realism
- "studio lighting, white background" → clean product shots
- "editorial photography, Vogue style" → fashion/portrait quality
- "cinematic, anamorphic lens flare" → movie poster feel
For logos, posters and anything with text: switch to ChatGPT or Gemini, which now render text accurately (see our test), or Adobe Firefly (trained on licensed content for commercial use).
Pricing: Midjourney is subscription-only; check its site for current plans and any trial offers.