Claude Code Tools

Cache and Context Prompts: 6 for Long Sessions That Cost Too Much

Astra can now change reasoning effort without invalidating the prompt cache, and Fable 5.1 prices cache reads at a fraction of fresh input. Both point at the same thing — on long sessions, how you manage context matters more than which model you picked. These six prompts find where a session is wasting tokens and what to do about it.

⌨️ 6 prompts 🕐 Updated Sep 6, 2026

💡 How to use these prompts: Replace everything in [BRACKETS] with your specific details before sending. Click Copy to copy any prompt to your clipboard instantly.

Find what is breaking my cache

Cache invalidation is almost always something small placed in the wrong position.

Here is how my requests are structured: [DESCRIBE OR PASTE]

Tell me:
- Which parts stay identical between calls and which change
- Where the changing parts sit relative to the stable ones
- What is invalidating the cache that does not need to
- How I would restructure so the stable prefix is longer

Cache breaks are usually a formatting decision, not a content one.

Work out what actually needs to be in context

Passing everything every step is the single biggest waste in most agent loops.

My agent receives this on every step: [DESCRIBE]

For each item tell me whether it is needed every step, only some steps, or could be fetched on demand.

Then estimate roughly what proportion of my context is carried unnecessarily.

Rank changes by tokens saved against risk of breaking something.

Decide where to spend reasoning effort

Now that effort can vary without losing the cache, this is the cheapest optimisation available.

Here are the steps in my task: [LIST]

For each, tell me whether it needs deep reasoning or whether shallow would do, and what specifically would fail if I used less.

Be concrete about the failure. I want to run low effort by default and raise it only where it earns the cost.

Compress a session without losing the thread

Long conversations re-send accumulated history every turn. Restarting from a summary resets that cost.

We are partway through a long task. Write a handover containing:
- What has been done and decided
- Why each decision was made
- What remains
- Anything about this context worth carrying forward

Aim for something a fresh session could start from cold. Leave out anything I could re-derive cheaply.

Estimate a session before running it

Estimating first is the difference between a budget and a surprise.

I want to run this task: [DESCRIBE]

Before starting, estimate:
- Roughly how many steps this takes
- How much context each step carries
- Where the cost concentrates
- Whether any part should run on a cheaper model

If it looks expensive, propose a cheaper structure before I commit.

Audit a month of spend for one change

The last line keeps it honest. Not every workload has waste in it.

Here is my usage breakdown: [PASTE]

Tell me the single change that would cut the most cost with the least risk, and what I should measure to confirm it worked.

One change. Not a list. If the honest answer is that my usage is already efficient, say that.