AI Guides › Workbench
By Nigel Guy · 7 min read
Most people deal with a foggy Claude chat in one of two ways. They push on while the answers get vaguer, or they open a new chat and retype what they remember, which is never everything. Both feel fine at the time. The first quietly lowers quality, and the second quietly drops decisions you made three hours ago and have since forgotten you made.
The rule: never leave a long chat until its decisions, current state and next steps are written into one file, and never start the new chat until that file is the first thing in it.
Claude can only work from what is in its context window, the text it can see when it writes each reply. Anthropic's developer documentation is blunt that more is not automatically better: as the token count grows, accuracy and recall degrade, a phenomenon it calls context rot. A bigger window delays the problem; it does not remove it.
Two vendor facts are worth knowing, and both change often, so check the current pages:
A handoff file is you choosing what survives.
| Tool | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Handoff prompt (below) | Claude writes a self-contained brief from the old chat | Included in your plan; uses a little of your usage allowance | Any chat, any plan | You must read the output; Claude can miss things |
| Pickup prompt (below) | Starts the new chat so it confirms the brief before working | Included in your plan | Every handoff | Skipping the confirmation step is where errors slip in |
| Project knowledge file | Stores the handoff where every chat in a Project can use it | Projects are available on free accounts too (free accounts: up to five projects, per the help centre) | Work that runs for weeks | Files go stale; you must replace them |
Claude Code /compact with instructions |
Summarises a coding session while keeping what you name | Uses your plan allowance or API tokens; check your own pricing page, in £ at checkout | Coding sessions | Compacting a huge context is itself a large request |
Claude Code /clear |
Starts fresh at no cost | Free | Unrelated next task | Wipes context; pair with a handoff file if the task continues |
The best moment is when a piece of work reaches a natural stop, or when replies start forgetting earlier decisions, repeating themselves or contradicting you. Do it in the old chat, because that is the only place the full detail exists.
Paste this into the old chat and fill in the brackets.
You are my project archivist. This chat is about to end, and a different Claude session, with no memory of it, will continue the work. Write a handoff file that session can act on without asking me to repeat anything.
Project: [PROJECT_NAME_AND_ONE_LINE_GOAL]
Who the work is for: [AUDIENCE_OR_CLIENT]
What "done" looks like: [DEFINITION_OF_DONE]
Go through this whole conversation and produce these sections, in this order:
1. Goal and audience, in three sentences or fewer.
2. Decisions made. One line each, with the reason, and mark any that I only half agreed to as "tentative".
3. Rejected options. What we ruled out and why, so it is not suggested again.
4. Current state. What exists now: drafts, files, numbers, names, versions. Quote key wording or figures exactly rather than paraphrasing.
5. Open questions. Things still undecided, and who needs to decide.
6. Constraints and preferences. Tone, format, length, things I dislike, facts that must not change.
7. Next steps. Numbered, with the very first action marked clearly.
Rules: use only what is in this conversation. If something important is unclear or missing, list it under open questions instead of guessing. Do not invent numbers, sources or decisions. Keep it under [WORD_LIMIT, e.g. 800] words, in plain text I can paste into a new chat.
Before you answer, check that every decision I stated appears in section 2 or 3, and that someone who has never seen this chat could start step 1 of the next steps straight away. Fix any gap, then give me the file.
Fill in: the project name, audience, what "done" means and a word limit.
This is the step people skip. Read the file against your own memory. Check three places in particular: the rejected options (Claude often omits these), anything marked tentative, and every number or name. If something is missing, reply in the old chat with "Add this: ..." and ask for the full file again. Fixing it now costs one message; fixing it after the old chat is gone costs the original reasoning.
Open a fresh chat, ideally with the same model, and paste this with the file underneath.
You are taking over a project mid-way. Below is a handoff file written by a previous session. Treat it as the source of truth, but do not assume it is perfect.
Step 1: Summarise the goal, the three most important decisions, and the first next step, in under 120 words.
Step 2: List anything in the file that is unclear, contradictory or missing something you would need.
Step 3: Stop and wait for me to confirm. Do not start the work yet.
Rules: do not reopen anything listed under rejected options unless I ask. Do not invent details to fill gaps; ask me.
Before you answer, check that your summary matches the file and that you have not added anything it does not say.
Handoff file:
[PASTE_HANDOFF_FILE_HERE]
Fill in: paste the file where the placeholder is. The confirmation stop is the check that it worked: if the summary is wrong, the handoff was wrong, and you have caught it before any work was built on it.
If the work runs over days, make the handoff a living document. In a Claude Project, you can add files to the project knowledge and write project instructions, per the help centre. Save the corrected handoff as a file there, and replace it at each stopping point rather than adding a second version. Name files with a date, such as handoff-2026-10-04, so you can tell which is current. I could not confirm from the help-centre page I checked whether Claude can read other chats in a Project, so do not rely on that; rely on the file.
In Claude Code, Anthropic's cost documentation lists /compact with focus instructions (its example is /compact Focus on code samples and API usage), /clear to start fresh between unrelated tasks, /rename before clearing and /resume to return, and a "Compact instructions" section in CLAUDE.md. Use /compact when you want to carry on in the same session, and the handoff prompt plus /clear when you want a clean slate with a record you wrote yourself. The docs also note that compacting a large context is itself a large request, while /clear costs nothing.
/compact with named priorities./clear. No handoff needed.