How to Milk Claude to the Last Drop
There's a specific moment I want to talk about.
You're deep in something good. Claude is finally producing the thing you actually wanted, you've got momentum, and then the little bar fills up and the session's done. And your first instinct is to be annoyed at the plan.
Sometimes that's fair. Usually it isn't. Usually you burned a third of that session re-explaining a project Claude could have already known about, uploading the same PDF for the fourth time, and sending seven messages that could have been one.
Soo, this is a guide to get the thing properly. Two kinds of achiveings, really: more quality per message, and more result per limit. They're different problems and they have different fixes.
Almost most of thing here got from offical documentation, which is genuinely good and which approximately nobody reads(hah). I've linked all of it at the bottom.
Let's start with the part that's going to annoy some of you.
In July 2026, Anthropic published something that should have been bigger news than it was. While tuning Claude Code for the Claude 5 generation of models, they deleted over 80% of its system prompt and saw no measurable loss on their coding evaluations.
Eighty percent. Gone. Nothing broke.
Their internal read on why is blunt: they'd been over-constraining the model. When they went through transcripts of their own usage, they found the system prompt, the skills, and the user's actual request all shouting slightly different things at once. One layer says leave documentation where appropriate, another says never write comments, and now Claude has to spend effort untangling which of its instructions you actually meant instead of doing your task.
Sound like your prompts? well,
Here's the "then and now" list, straight from that post, translated into things you might be doing in a chat window right now.
Then: pile on rules. Now: let it use judgement. Older models needed hard guardrails because their defaults were bad. Newer ones have better judgement(kinda blocking things), and a rule that's right 80% of the time actively hurts you the other 20%. Anthropic replaced rigid comment rules in Claude Code with a single line about matching the surrounding code's style. That's the shape you want: describe the target, don't legislate every case.
Then: stack up examples. Now: design the interface. This one surprised me. Anthropic found that examples can constrain the newest models into a narrower exploration space than you wanted. Examples are still useful when format is easier to show than describe, but three examples aren't three times better than one, and a sloppy example is worse than none. These models read your examples very closely, including the parts you didn't mean to be instructive.
Then: put everything upfront. Now: progressive disclosure. The instinct to build one giant mega-prompt containing every rule you might ever need is understandable and wrong. Load context when it's relevant, not preemptively. In chat, this means a lean project instruction plus documents Claude can reach for, not a 900-word wall of preferences it has to reconcile on every single turn.
Then: repeat yourself. Now: say it once. Older models paid uneven attention across a long context, so people learned to repeat the important bits at the top and the bottom. That's largely obsolete, and repetition now mostly just costs you tokens.
And two crowd favorites that Anthropic's own prompt engineering guide now files under "techniques you might have heard about":
XML tags used to be the recommended way to structure anything complicated. Modern models understand structure fine without them. Still handy when you're mixing several kinds of content and want unambiguous boundaries. No longer a requirement.
Heavy role prompting is the other one, and it's the one I want you to actually let go of. Anthropic's specific warning is against over-constraining the role. Their example of what not to do: "You are a world-renowned expert who only speaks in technical jargon and never makes mistakes."
You know that 400-word preamble you paste at the top of every request? You are an elite senior brand strategist with 20 years of experience at top agencies, known for your razor-sharp insight and—
Delete it. Replace it with the perspective you actually want: "Analyze this positioning with a focus on which claims a skeptical buyer would push back on." That's the whole trick. Say what lens you want applied instead of casting a character.
Fine, so what survived? Anthropic's core techniques, all of which are boring and all of which work.
Be explicit, and ask for depth on purpose. Their example is a good one. "Create an analytics dashboard" gets you a shrug. "Create an analytics dashboard. Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation" gets you something you can use. Modern Claude doesn't pad by default, so if you want comprehensive, request comprehensive.
Explain why, not just what. This is the highest-leverage habit in the entire guide and it takes eight extra words. Compare:
"NEVER use bullet points"
"I prefer responses in natural paragraph form rather than bullet points because I find flowing prose easier to read and more conversational. Bullet points feel too formal for my casual learning style."
The first is a rule that will be applied literally, including in the one case where a list was genuinely correct. The second gives Claude the reason, so it can extend your preference to situations you never anticipated. Rules break at the edges. Reasons generalize.
Be specific about the constraints that matter. "Create a Mediterranean meal plan" versus a plan for pre-diabetic management at 1,800 calories with low-glycemic emphasis, broken into three meals and a snack with nutritional breakdowns. Same request, wildly different output, about fifteen seconds of extra typing.
Give it permission to not know. Add "if the data is insufficient to draw conclusions, say so rather than speculating." Anthropic recommends this explicitly as a hallucination reducer, and it costs you one sentence. This is the single cheapest reliability upgrade available.
Say what to do, not what to avoid. "Do not use markdown" performs worse than "write in flowing prose paragraphs." Negative instructions leave the target undefined. Related trick from the same doc: your prompt's own style influences the response style. Write your prompt in dense markdown and you'll get dense markdown back. Write it conversationally and watch what happens.
protect your intellectual . we don't know is our data is using to train things or not, If you have some kind of million worth ideas , writings, blueprints, etc better not feed them or ask via uploading to non local AIs.
Now the part you came for.
First, know what's actually draining the tank. Per upport docs, your usage is affected by message length, file attachment size, current conversation length, tool usage like research and web search, model choice, effort level, and artifact creation.
Read that list again and notice item three. A long conversation costs more per message than a short one, because the whole history rides along on every turn. That forty-turn epic you've been nursing since Tuesday is expensive now in a way it wasn't on turn three.
Then here's the lever nobody pulls.
Projects are a caching system and you're not using them that way
Straight from the docs: content in projects is cached and doesn't count against your limits when reused. Upload your reference documents once, and subsequent questions about them only charge you for the new parts.
Sit with the implications. If you're uploading the same style guide, the same spec, the same three research PDFs into fresh chats over and over, you are paying full freight every single time for material Claude could be holding for free. Put it in a project's knowledge section once and ask fifty questions against it.
load all your reference materials into a project, then ask as many questions as you want while burning far less than you would re-uploading each time. There's also a RAG mode for projects when your knowledge base outgrows the normal capacity.
The docs also note that similar prompts you use frequently are partially cached. So that daily "summarize this and pull the action items" phrasing you've settled into? Consistency is quietly cheaper than creative rephrasing.
Batch like you're doing groceries
group related tasks into one message. Five separate math problems in five messages costs five messages. Five problems in one message costs one.
Apply this everywhere. Send the whole document for editing, not three sections in sequence. Give your full coding environment in the opening message rather than dribbling out context as Claude asks for it. Every clarifying question Claude has to ask you is a round trip you paid for and didn't need.
Read your own meter
Settings → Usage shows progress bars for both your five-hour session limit and your weekly limits, plus time remaining in the current session.
Which leads to the obvious move nobody makes: stop running your heaviest model on trivial tasks.
Know when to abandon a chat
Two failure modes here and you want to avoid both.
Starting fresh for every tiny follow-up throws away context you already paid to establish. But nursing one enormous conversation means every new message drags an increasingly heavy history behind it.
Rough heuristic: when the thread has become a junk drawer, when you've pivoted topics twice and half the history is dead ends, that's the moment. Ask Claude to summarize the current state and decisions, start clean, paste the summary. You just compressed forty turns into a paragraph.
There are three separate personalization layers in Claude and most people use zero of them.
Profile instructions are account-wide, applied to everything. This is where your durable stuff goes: what you do, who you write for, how you like things formatted, what you never want. Set once, benefits every conversation forever.
Project instructions apply only inside that project. This is where the context that would be wrong elsewhere goes.
Styles handle voice and format specifically, so you can switch tone without rewriting your instructions.
Then the memory layer. On paid plans, Claude can search your past chats — it's a retrieval system, it shows up as tool calls, and it's scoped either to all your non-project chats or to within a single project. Separately, Claude can generate memory from your chat history, saving relevant facts as you work. Projects keep their own memory separate from your non-project chats, and on Team and Enterprise you can move conversations in and out to control exactly what's included.
Turn these on. Then go look at what's actually in there and prune it, because a memory full of stale facts from an abandoned project is worse than no memory.
The one prompt that builds your instructions for you
This is the best trick in the article and it takes thirty seconds.
After a working session where you've corrected Claude a few times, send this:
"Look back at what I kept versus what I rewrote in this conversation. What pattern do you see in my corrections? Draft two or three specific instructions I could add to my profile or project settings that would prevent this friction next time."
You're mining your own edits for preferences you've never managed to articulate. I've gotten instructions out of this that I'd never have written myself, because I didn't consciously know I had the preference until it was named back to me.
Make it interview you. For anything vague, open with: "Before you start, ask me the three questions whose answers would most change your approach." Two minutes of questions beats four rounds of revision, and it costs a fraction of the messages.
Ask for the rubric before the work. "What are the five criteria a great version of this would meet?" Then have it produce the thing. Then have it grade itself against its own criteria and fix the gaps. Anthropic's guide calls this chaining — draft, critique, revise — and the accuracy gain is real. Three focused prompts consistently beat one prompt trying to do everything.
Edit, don't argue. When a response misses, resist replying "no, not like that." Go back and edit your original message with what you now know you should have said. You get a clean attempt instead of one contaminated by the failed attempt sitting in the history.
Force clean output. True prefilling is an API feature, but the docs give the chat approximation: "Output only valid JSON with no preamble. Begin your response with an opening brace." Works for any format where you want the content and not the conversational wrapper.
Ask what it needs. "What additional context would let you do this significantly better?" Sometimes the answer is nothing. Often it's a document you forgot you had.
Said plainly, so you can stop:
Threatening it. Tipping it. Telling it your career depends on this. Telling it your grandmother is dying. The elaborate emotional theater that circulated in 2023 is not doing what the screenshots claimed it was doing.
Stacking five personas. Repeating the same instruction in caps three times. Reciting the size of the context window at it like a spell. Pasting a 700-word preamble before a request that was one sentence long.
claude (anthorpic) own list of mistakes to avoid is refreshingly short: don't over-engineer, don't use every technique at once, don't skip the basics, don't assume mind-reading, and don't lean on outdated techniques. Longer is not better. The best prompt is the one that gets the result reliably with the minimum necessary structure.
The whole thing, compressed
Set up once:
- Profile instructions with your durable preferences, written as reasons rather than rules
- A project for each recurring body of work, core documents loaded into knowledge
- Memory and past-chat search enabled, then reviewed monthly
- Model choice matched to task weight, checked against Settings → Usage
Every time:
- One message, batched, with the full context and constraints in it
- Say what you want, why it matters, and how deep to go
- Permission to say "I don't know"
- Edit the prompt instead of arguing with the output
Never:
- Re-upload a document that could live in a project
- Write a persona preamble longer than the actual request
- Nurse a dead conversation because you don't want to lose the history
links
- Best practices for prompt engineering — Anthropic
- The new rules of context engineering for Claude 5 generation models — Anthropic, July 2026
- Usage limit best practices — Claude Help Center
- Understanding Claude's personalization features — Claude Help Center
- Use Claude's chat search and memory to build on previous context — Claude Help Center
- Prompting best practices — Claude Platform Docs
AI related things changes and limits change often. If something here doesn't match what you're seeing, the support center is more current than I am. just nothing is permanent
Comments
Post a Comment