LovableUse Chat Mode for planning, Build Mode for assembly
Chat Mode costs ~1 credit per message; Build Mode is usage-based and far more expensive. Plan your architecture in Chat first, then send one polished Build prompt.
Example
Chat: 'Outline the data model for a tasks app with teams.' → Build: 'Implement the schema, components, and routes from the plan above.'
LovableBatch related changes
Five tiny Build Mode messages cost much more than one batched message. Group UI tweaks, copy fixes, and small bugs into a single prompt.
Example
Instead of 'fix header', then 'fix footer', send: 'Fix header padding, footer color, and the hero button alignment.'
LovableUse templates and existing components
Generating from scratch burns more tokens than tweaking existing scaffolds. Start from a template close to your goal.
Example
Pick a SaaS landing template, then ask only for what's different.
ClaudeKeep your SKILL.md short
In Claude Code, every byte in SKILL.md is re-sent each turn. Trim aggressively — link out to docs instead of pasting them.
Example
Replace a 200-line style guide with: 'See style.md for full rules.'
ClaudeClose unused files between turns
Open files inflate context. Close anything you're not actively editing before sending the next message.
Example
After finishing the API layer, close those files before starting on the UI.
ClaudeStart fresh chats for new tasks
Carrying a 50-message thread into a new feature means you pay for every previous token. Fork a new chat.
Example
Done with auth? Start a new conversation for the billing flow.
CursorPrefer /compose over Cmd+K on large files
Cmd+K sends the whole file as context. /compose is targeted and uses far fewer tokens for the same edit.
Example
Use /compose 'add error handling to handleSubmit' instead of Cmd+K on a 500-line component.
CursorPin only what's needed
Pinned files are sent every turn. Unpin once they're no longer relevant.
Example
Pin auth.ts while building login, unpin once shipped.
ChatGPTUse GPT-4o-mini for simple tasks
Drafting copy, summarizing logs, formatting JSON — none of these need GPT-4o or o1. Mini is 10–30× cheaper.
Example
Use mini for: 'Rewrite this paragraph in 30 words.' Use the heavy model for: 'Design a sharded queue system.'
ChatGPTSet custom instructions to skip preamble
If GPT keeps adding 'Sure! Here's what you asked for:' you're paying for those tokens. Tell it to skip pleasantries.
Example
Custom instruction: 'Respond directly without restating the question or adding closing remarks.'
BoltIterate with smaller diffs
Bolt re-bundles your project on each change. Small, focused asks ship faster and cheaper than 'redesign everything'.
Example
Ask for one component at a time, not 'a full marketing site'.
ReplitUse Checkpoints to roll back, not re-prompt
Replit Agent charges for each fix attempt. When something breaks, restore to the last good Checkpoint instead of asking the Agent to undo its own changes.
Example
Agent broke routing? Click the previous Checkpoint → reapply your last working prompt.
ReplitKeep the file tree small
Agent re-reads project context each turn. Delete unused scaffold files and demo folders early — every kept file inflates the per-message cost.
Example
After picking a template, prune the example pages you won't use before issuing your first real prompt.
WindsurfUse Cascade Write for edits, Chat for planning
Cascade's Write mode applies surgical, scoped diffs; Chat is cheaper for thinking out loud. Don't use Write to brainstorm.
Example
Chat: 'How should I structure the auth flow?' → Write: 'Add a useAuth hook to src/hooks based on that plan.'
WindsurfTrim your .windsurfrules file
Every line in .windsurfrules is sent on every turn. Treat it like SKILL.md — link to docs instead of pasting them.
Example
Replace a 100-line coding standard with: 'Follow conventions in docs/style.md.'
Claude CodeUse /clear between unrelated tasks
Claude Code keeps the full conversation in context until you clear it. Switching from auth to billing? Run /clear first to drop the auth tokens you no longer need.
Example
Finished the login flow → /clear → start the Stripe integration with a fresh context window.
Claude CodePrefer Read+Edit over pasting files
Pasting a file into your prompt costs tokens twice (your message + Claude's working copy). Let Claude Read it directly and Edit in place.
Example
Say 'Read src/api/users.ts and add input validation' instead of pasting the file contents.
GitHub CopilotInline completions for boilerplate, Chat for logic
Tab-completions are essentially free per keystroke; Chat invocations cost a request. Use inline for getters, types, and obvious patterns; save Chat for non-obvious logic.
Example
Let inline finish a useState hook. Use Chat only when you need 'why is this useEffect re-running?'
GitHub CopilotScope @workspace with file globs
@workspace by default searches the whole repo. Narrow it with explicit paths so Copilot doesn't pay to index irrelevant folders.
Example
Use '@workspace /src/api how is auth handled?' instead of '@workspace how is auth handled?'
GeminiFlash for summarization, Pro for reasoning
Gemini Flash is roughly 10× cheaper than Pro and handles summarization, classification, and extraction just as well. Reserve Pro for genuine multi-step reasoning.
Example
Flash: 'Summarize these 20 support tickets into 5 themes.' Pro: 'Design a multi-tenant data model for this spec.'
GeminiUpload reference docs once
Gemini's long context lets you attach a single PDF or doc and reference it across many prompts — far cheaper than pasting excerpts into each message.
Example
Upload your API spec once, then ask 'Generate a client for the /orders endpoint per the attached spec.'