Token costs are the number one complaint in AI coding right now. Most of the damage comes from a few default habits that are easy to fix.

I tested every optimization I could find this month. These 7 made the biggest difference in Claude Code:

1. Clear context between tasks. Type /clear when you switch tasks. Every new message re-sends your full conversation history as input tokens. That debugging session from an hour ago is still inflating every prompt you send. A fresh start costs nothing. Carrying stale context costs you on every turn.

2. Compact at 60%, not 95%. Claude auto-compacts near 95% context capacity. By then, output quality has already degraded. Run /compact focus on [current task] at 60% yourself. You get a cleaner summary and stay in the range where the model still performs well.

3. Match the model to the task. Opus for complex reasoning. Sonnet for routine code. Haiku for simple lookups and formatting. Most tasks don't need the most expensive model. One team documented a 72% cost reduction just from model switching and prompt caching over three months.

4. Offload heavy reads to subagents. A 10,000-line log file that Claude reads early in a session stays in context for every message after it. Instead of reading it in your main session, spin up a subagent. It reads in isolated context and returns only the findings. Your main window stays clean.

5. Build deterministic tools that cost zero tokens to run. Not everything needs an LLM call. Data formatting, file moves, test runners, API calls with known inputs. Write these as regular scripts. The LLM orchestrates. Deterministic code executes. The scripts run for free, every time, with predictable output.

6. Keep CLAUDE. md lean. It loads into every session before anything else. A 5,000-token CLAUDE. md costs 5,000 tokens before you've typed a word. Every turn. Every session. Keep it under 200 lines. Move project-specific context into scoped markdown files that only load when relevant.

7. Run /usage before starting a new task. Don't wait until you notice the model making mistakes it wouldn't have made 20 minutes ago. Check /usage, see where you stand, and decide whether to /compact or /clear before committing to the next chunk of work. Most of these take less than a minute to build into your workflow.

The compounding effect is what matters. Each one shaves a layer. Run all 7 together and the same session that used to drain your budget in an hour lasts the full workday.

Robert Cialdini spent decades studying persuasion.

His 7 principles explain why people say yes.

I turned each one into a prompt that rewrites your copy to use it.

Here are all 7. Steal them: 1. Reciprocity: The Give-First Rewrite
(Cialdini's principle of reciprocity, from Influence)

Your job: rewrite the copy I give you so it gives something useful before it asks for anything.

What to do:
1. Find the ask in my copy and hold it back until value comes first.
2. Add one helpful thing up front: a tip, a resource, or a quick win the reader keeps even if they say no.
3. Make the give specific and real, not a vague promise.

Rules:
The give has to be something I can deliver.
Keep my voice and my facts.
One give is enough. Do not pile on.

Output: a rewritten version of my copy, plus a one-line note on the single change doing the most work.

🚨 BREAKING: Claude has a feature called Red Team Mode.

You can use it to attack your own business the way a competitor, investor, or angry customer would, and fix the weak spots before they become real problems.

Here are 7 prompts to access it: 👇 1. The Competitor War Game

Prompt: "You are my smartest, most ruthless competitor. I'm going to describe my business model, pricing, and target customer.

Your job: build the exact strategy you'd use to steal my customers, undercut my positioning, or make my product irrelevant within 12 months.

Be specific. Name the moves. Show me the timeline.

Here's my business:

[DESCRIBE YOUR BUSINESS MODEL, PRICING, TARGET CUSTOMER, AND KEY DIFFERENTIATOR]"

🚨 BREAKING: Every major LLM has a hidden Anti-Slop Response Mode.

You can use it to stop getting watered-down, over-formatted, sycophantic answers that all sound the same.

Here are 7 custom instructions to activate it: 👇 1. Strip fillers.

Add to custom instructions:

"Never open with filler like 'Great question!' or 'Absolutely!' or 'That's a really interesting point.'

Start with the answer. If you don't know, say so. If you disagree with my premise, say that instead."