# Working Lean > Gets faster, more accurate answers by keeping each conversation to one matter and sending the material that bears on it without the material that does not. Use when someone is about to attach a mailbox or file export, a whole drive or client folder, or many documents at once; when a conversation has run long, drifted across unrelated matters, or started repeating itself, losing detail or contradicting something said earlier; when someone has reached or is approaching a usage limit, or asks how to make their allowance last; and when choosing which folder to point a Cowork task at. Covers what to attach and what to leave out, why completeness beats economy on a lease or a set of accounts, and when to start a fresh conversation instead of continuing. - Skill: `randommonicle/working-lean` (Agent Skill) - Install (CLI): `npx skillmds@latest add randommonicle/working-lean` - Raw SKILL.md: https://api.skillmd.com/api/skills/randommonicle/working-lean/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: randommonicle (https://skillmd.com/u/randommonicle) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/randommonicle/working-lean --- # Working lean Getting a better answer and using less of your allowance are usually the same action. This is not a thrift measure, and it is not about sending less than a question needs. ## The one thing worth knowing Claude does not remember your conversation between messages the way a colleague does. Every time you send a message, **the entire conversation so far is sent again**. Every earlier question, every answer, and the full text of everything you have attached goes up with it. So a short question at 4pm in a conversation you started at 9am is not a short question. It carries the whole morning with it. Two things follow, and everything below is one of them: 1. **What you attach, you pay for on every later message**, not once. 2. **Irrelevant material is not harmless.** It competes for attention with the part that matters. Independent testing across eighteen different AI models found that adding even one piece of irrelevant material to what the model was given measurably reduced accuracy. A long conversation carrying four unrelated matters is not a tidy filing system. It is four distractions. ## One matter, one conversation Start a new chat when the subject changes. A conversation about a section 20 consultation at Property A should not later be asked about a service charge query at Property B. The signs you have stayed too long in one conversation: - It starts repeating things it already told you. - It loses a detail you gave it earlier. - It mixes up two properties, two leaseholders, or two matters. - It contradicts something it said higher up. None of these mean it has gone wrong. They usually mean the conversation has got too big to hold cleanly. Start a fresh one, paste in the two or three facts that still matter, and carry on. That is generally quicker than correcting it. ## Send what the question needs, and no more Completeness beats economy every time. Sending too little is a worse mistake than sending too much, and this section is about the difference between material that bears on a question and material that does not. It is not permission to withhold. **A lease is one document and must be read as one.** A repairing covenant cannot be read on its own: its meaning depends on the definitions, the extent of the demise, the landlord's reciprocal covenants, the service charge schedule, any deed of variation, and often the insurance provisions. If the question is a lease interpretation question, **send the whole lease and every deed that varies it**. The same applies to a set of accounts, a section 20 consultation trail, or anything where a later clause can change the meaning of an earlier one. An answer drawn from a partial lease is worth nothing, and may be worse than nothing. What this section is actually about is material with no bearing on the question at all: - **A mailbox export, a full drive listing or an entire client folder is never the right attachment.** It is slow, it is less accurate, and our standing rule is to work only from a folder scoped to the task in hand. - **Ten properties' worth of documents when the question concerns one.** Send that one's. - **Say where the answer is likely to be** when you send something long: "the repairing obligations are at clauses 4 and 7" tells Claude where to start. That is a pointer, not a restriction, and it does not stop the rest being read. If you are unsure whether something bears on the question, send it. The cost of the extra pages is small. The cost of advice given on a document that was missing its definitions is not. ## Put your question at the end If you are sending a long document or a long set of notes, put the context first and your actual question last. Google publishes this as an accuracy finding: performance is better when the question comes after the material. (Anthropic recommends the same running order for a different, technical reason, so one layout suits both.) Say what you want out, too. "A two paragraph summary", "a table of the dates", "a draft letter to the RMC directors". An unstated format gets you a guess, and then a second attempt. ## In Cowork: choose the folder before you start Cowork works from a folder, and the folder you choose shapes both the quality of the work and the size of the job. - **Point it at the narrowest folder that holds what the task needs**, not at a client folder, a portfolio folder, a mailbox export or a drive. This is an ASH rule, and it is also simply how you get a good answer. - **If the material is spread across several folders and the job is substantial, gather it into one working folder first.** Skip that when you are answering something quickly, and just point at the narrowest folder that covers it. - **Say what is in the folder** when you start. "This folder holds the last three years of service charge accounts for one block" orients the work immediately. ## Why you hit a limit, and what actually helps Usage limits are reached by total volume across a rolling window, and on Team and Enterprise plans the same allowance covers chat, Cowork and Claude Code together. It is often not the number of questions asked but a few very large conversations, each re-sending everything in them on every message. What helps, roughly in order: 1. **Start a new conversation between unrelated matters.** Usually the biggest lever, and it costs nothing. 2. **Leave out material that has no bearing on the question** (which is not the same as leaving out material the question needs). 3. **Do not leave a large conversation open and return to it hours later** for an unrelated question. Start a new one. 4. **Ask for what you want in one go** rather than in eight small corrections, each of which re-sends the whole thing. Asking Claude to be brief helps much less than people expect. The reply is a small part of the total; the conversation and the attachments are the large part. ## Sensitive material: what must never go in, and what to do about the rest Separate from anything about efficiency. There are two rules here doing two different jobs, and the difference matters. **Rule 1: never send it deliberately. This one is absolute.** - **No bank account numbers, sort codes or payment card details** pasted in, and no document attached *in order to* work on those details. - **No payroll, HR or salary data.** - **Nothing subject to legal professional privilege.** **Rule 2: a document you genuinely need does not become unusable because a bank detail appears in it somewhere.** Invoices, remittance advices and service charge packs often carry one. You have two routes, and either is acceptable: - **Redact it and send the document**, where that is quick to do. - **Where redacting is not practical**, such as a long statement or a scanned pack, send it and say so in your first message: *"this pack contains bank details; do not reproduce them or refer to them."* Then note on the matter file that you did. Claude will not repeat them back or carry them into anything it drafts. Be clear about what route 2 does and does not achieve. It stops those details reaching a letter, a report or a summary. It does not unsend them: they are still in that conversation. So it is a sensible working compromise for material you actually need, and it is not a reason to relax about what you attach. Prefer redaction where you reasonably can. **If details go in that should not have**, whether pasted by mistake or inside a document nobody checked, that is a data incident. Stop, do not carry on in that conversation, and tell **Simon McKeag, Data Protection Manager** straight away. Telling Claude does not deal with it. Remember which party is which, because it changes the advice: our **clients** are the Resident Management Companies, Right to Manage companies, freeholders and landlord clients who instruct us. **Leaseholders** and **tenants** are not clients, and are not the same as each other. Anything Claude drafts that is going out to a client, a leaseholder, a tenant, a contractor or a tribunal needs a **reliability decision** recorded before it is issued. That means you, as the person responsible for the matter, have read the draft, satisfied yourself it is correct, and noted on the matter file that you checked it and are content for it to go. It is your professional judgement on the record, not Claude's. If you are unsure what to record or where, ask Ben Graham before issuing. ## Questions AI use, including this skill: **Ben Graham, AI Lead**. Personal data, and anything you are unsure about attaching: **Simon McKeag, Data Protection Manager**.