What You Need to Know
Complex work gets Claude's best output when split into pieces it can do properly, one at a time. That splitting is task decomposition. Connected parts feed each other: a later part needs an earlier part's output or decisions. Independent parts stand alone.
A single well-formed prompt handles a lot of work, but it has a ceiling. Cram a project's worth of asks into one message and the result is shallow and uneven: some parts done well, others skimmed. Take a supermarket manager writing a quarterly customer-feedback report for head office. It needs an outline, sections in one voice and an action list. She also needs a rival-store summary, a press release and a budget table. The Help Center's remedy list for an unhelpful answer includes breaking complex requests into substeps. It also says a prompt should contain as much context as possible.
Four Techniques, Lightest to Heaviest
Numbered steps inside one prompt. For connected parts, list them explicitly. The manager writes: first summarise the feedback themes, then rank them by frequency, then draft one improvement action per theme. Claude follows an explicit sequence far more reliably than a run-on paragraph of asks. Anthropic's guidance is to use numbered sequential steps when order or completeness matters, and to be specific about output format and constraints.
Staged prompting in one conversation. For a deliverable built in layers, ask for the outline, review and adjust it, then draft section by section. You inspect each stage's output before the next stage runs. That approved output becomes the next stage's foundation, so course corrections happen while still cheap. The manager fixes her outline before any section exists.
Plan-first prompting. When the right structure is not obvious, ask Claude to propose an approach first. For example: before writing anything, list the sections you would include and what each covers. Reviewing a plan takes seconds. Reworking a wrong 2,000-word draft does not.
Separate conversations per sub-task. Genuinely independent pieces, such as a competitor summary, a press release or a budget table, go in separate chats. That keeps each context clean and focused. The manager's three side jobs get a chat each.
Choosing the Technique
Connected parts belong in one conversation, staged, so later stages inherit the approved earlier ones. Independent deliverables belong in separate chats. When the structure is unclear, ask for the plan first. A mismatch fails one of two ways: a bloated mega-prompt asking for a finished 20-page product in one pass, or connected work fragmented across chats. Each new chat has then lost the outline and tone decisions the last one built. Report sections split across chats would lose her outline and tone.
Carrying Context Forward
A new conversation does not automatically contain another conversation's working material. Claude does now remember across chats, as a set of individual topics covering your role, projects, preferences and ongoing work context. It is context about you, not a transcript. It does not carry your approved outline, tone guide or half-finished draft into a new chat. On paid plans you can ask Claude to search past chats and pull in relevant details. That is retrieval you must prompt for, not automatic carry-over.
The dependable pattern is explicit persistence. Paste the output into the new conversation, re-attach the file, or keep it in a Project's knowledge base. Projects are self-contained workspaces with their own chat histories and knowledge. Back on the report in a fresh chat, the manager pastes in her outline first. Carrying outputs forward is your job, not the product's.
In the exam
The exam tests matching a technique to a situation, not reciting the four. The keyed answer follows the three matching rules. Both failure modes appear as distractors. Where persistence should live is a Domain 3 and Domain 5 question.
Current state (September 2026)
Memory arrived on claude.ai on 11 September 2025 for Team and Enterprise plans, then for Pro and Max from 23 October 2025. As of 30 September 2026 it is on by default for Free, Pro and Max. As of that date, on Team and Enterprise an owner decides whether memory is available. It stays off for each member until that member turns it on. Chat search, asking Claude to reference past conversations, is available on paid plans. Neither feature changes the rule: memory saves context about you as individual topics, and chat search must be prompted.
Exam Traps
The do-everything mega-prompt, chosen to keep the whole deliverable in one place
Attention spreads unevenly across a project of asks in one message, so some sections come back strong and others skimmed, and corrections arrive only after the whole thing is wrong. Breaking complex requests into substeps is the documented remedy.
Fragmenting connected work across separate conversations
Each new chat lacks the outline, tone decisions and approved material the previous chat built, so the pieces arrive in different voices and need reconciling by hand. Separate chats are for independent deliverables, not connected ones.
Plan-first prompting offered where the structure is already known and agreed
Asking Claude to propose sections you have already settled is a small over-engineering detour: it adds a review round without adding information. Plan-first earns its place only when the right structure is genuinely unclear.
Practice Scenario
A programme director at a charity must produce a 20-page annual impact report: an overview, five sections drafted from the notes of five different programme teams, and one consistent voice throughout. The director wants to approve each part as it is produced, and never to explain settled decisions again to a new chat. How should the director structure the work with Claude?
Build Exercise
Stage a Multi-Part Deliverable
What you'll learn
- How to sort the parts of a deliverable into connected and independent
- How to assign the lightest technique that fits each part
- How to correct an outline before any drafting starts
- How to carry an approved output forward into later stages
- Choose a genuinely multi-part deliverable from your work: an annual report, a project plan, a course pack
Why: Decomposition only shows its value on work too big for one prompt; a small task will not exercise the techniques.
You should see: One deliverable named, with its parts listed as bullet points.
- Sort the parts into connected, meaning they feed each other, and independent, meaning they stand alone
Why: This sorting is the judgement the exam tests: connected parts want one staged conversation, independent parts want separate chats.
You should see: Each part labelled connected or independent, with one phrase saying why.
- Assign a technique to each part: numbered steps, staged conversation, plan-first, or a separate chat, with one sentence of justification each
Why: Matching technique to task is the objective itself; the justification keeps the choice honest.
You should see: A short list pairing every part with a technique and a reason.
- Run the first stage, the outline or the plan, and correct it before any drafting starts
Why: Course corrections at the outline stage cost seconds; the same corrections after 2,000 words cost an afternoon.
You should see: An approved outline you have actively edited, not just accepted.
- Each time a later stage needs an earlier output, note how you carried it forward: pasted in, re-attached, or stored in a Project
Why: A new conversation does not automatically contain the working material of an earlier one; explicit carry-forward is the reliable pattern.
You should see: A note per stage recording the carry-forward method you used.
Sources
- My prompt isn't giving me a helpful answer — Anthropic
- Prompting best practices — Anthropic
- How do usage and length limits work? — Anthropic
- Use Claude's chat search and memory to build on previous context — Anthropic
- Bringing memory to Claude — Anthropic
- What are projects? — Anthropic