Don't burn through Claude's limits work in context economy

Context economy rests on three pillars: cut the useless input, cut the useless history, cut the useless output. Here are the concrete moves that put them into practice.

1. 1

   Fix the original prompt, don't stack corrections

   Ten follow-up messages ("make it SEO", "add the callouts", "shorten it") all stay in memory, and you pay for them again with every reply. Better to rewrite the original prompt once, with every request folded in.

2. 2

   Open a new chat every 15-20 messages

   When the phase of the work changes, ask Claude for a summary of the essentials, copy it, and start fresh. That way you don't drag old, now-useless context forward.

3. 3

   Group related questions together

   One message with "summarize this, pull out the key points, suggest 3 titles" costs less than three separate messages that reread the same text every time.

4. 4

   Skip huge uploads you don't need

   Don't upload a 180-page PDF to analyze 3 of them. Extract the relevant sections or paste just the piece you need: every extra page is context you pay for.

5. 5

   Ask for the short version first

   Start with "give me a short outline" (max ~700 words), validate the direction, then expand only the right sections. You avoid huge outputs you might end up scrapping.

6. 6

   Clean up the context when you can

   Every so often, tell Claude: "ignore the naming part, just keep these 5 decisions." You bring the focus back where it belongs without opening a new chat.

Recurring context (brand voice, tone, SEO rules, templates, formatting standards) shouldn't be pasted into every chat. Put it in Projects: Anthropic puts it in cache, so you reuse it without paying for it again each time. Only new content counts against your limits.

Move long outputs (documents, code, structured content) into Artifacts instead: the chat stays clean and light instead of bloating with text inside the conversation flow.

Using the most powerful model for everything is the fastest way to hit your limits. Choose based on the task.

1. Opus

   complex reasoning, advanced coding, strategic analysis. The most powerful, but also the most expensive. Use it only when you truly need it.

2. Sonnet

   writing, editing, brainstorming, everyday operational work. It's your workhorse for most things.

3. Haiku

   tasks where speed matters: simple classifications, light passes, repeated micro-operations.

Rule of thumb: Sonnet as the default, Opus only when the problem earns it, Haiku for mechanical steps. If you use voice mode to brain-dump (great for thinking out loud), then ask Claude to extract the decisions and structure them, so you don't carry the raw monologue forward.

When a chat becomes long and expensive, don't keep pushing it. Ask Claude for a clean summary, copy it and start over in a new chat. Here's the ready-to-paste prompt.

I want to continue this work in a new chat without carrying over useless context.

Summarize only what's truly needed to continue:
- the goal of the work
- decisions already made
- important constraints
- required style
- materials used
- mistakes not to repeat
- next steps

Drop:
- failed attempts
- digressions
- repetitions
- messages that are now outdated
- details that are no longer relevant

Output in markdown, clear, concise, and ready to paste into a new chat.

You copy the markdown it gives you, open a new chat, paste it in and start clean: the goal, the decisions made and the next steps stay, the noise doesn't. Claude isn't magic, it's an amplifier: give it light, focused context, and the limits stop being a problem.