Episode 6

Long sessions

In this episode (≈ 20 min of hands-on work)

A fixed rhythm: one task per session, then /clear.
Subagents for research and a reviewer that checks your changes.
A continuation prompt so the next session continues without questions.

Your step

Paste this at the end of a session, before you type /clear. Claude saves the state into the three task files and hands you a short prompt for the next session.

We are ending this session. Before I type /clear:
1. Update tasks/current.md: status, decisions made, files changed, and under "Exact Next Action" the next step, precise enough that a fresh session can continue without asking me anything (file paths, function names, expected result).
2. Add to tasks/lessons.md every rule we learned from a mistake today, in the format "Do X / Don't do Y ... — learned after ...". If there were none, write nothing.
3. In tasks/todo.md, tick off only what is finished AND checked.
4. Tell me which test, lint and build commands you actually ran in this session and what the result was. Name explicitly anything you did not run.
5. Finally, give me a continuation prompt in a code block, ten lines at most. It starts with "Read tasks/current.md, tasks/lessons.md and tasks/todo.md." and then describes exactly the next step.

Files for this episode

prompt-continuation.mdThe continuation prompt, with an example result.Download ↓

Questions about this episode? Just email us: hello@ailoopwise.com

Step-by-step guide

What this episode is about

Working for hours in the same session burns tokens, and the AI gets less precise. The docs say it plainly: "performance degrades as it fills" (source). Short cycles work better: heavy sub-tasks go to subagents, a separate reviewer agent checks at the end, then you have Claude write a continuation prompt and wipe the history with /clear. The knowledge lives in your text files. The fresh session gets the new prompt and carries on with no baggage.

The wall: the workflow only clicks after a few weeks. Until then you'll lose sessions.

What you need

Your project with the blueprint's agents: /new-project creates researcher, implementer and reviewer in .claude/agents/.
The continuation prompt: examples/prompt-continuation.md
For reference: subagents https://code.claude.com/docs/en/sub-agents · context window https://code.claude.com/docs/en/context-window · reducing cost https://code.claude.com/docs/en/costs

Step by step

Look at your context. At any point in the session type:
```
/context
```
You see what takes how much space, including your CLAUDE.md.
One task per session. Start with the last session's continuation prompt, or with:
```
Read tasks/current.md and tasks/lessons.md. Summarize where we left off.
```
Offload research. Type @ and pick researcher (agent), or write the name directly:
```
@agent-researcher Find out how to build signed unsubscribe links in Node.js without storing tokens in the database. Return only a summary with sources, ten lines at most.
```
The subagent works in its own context window. Only the summary comes back, and your main window stays free.
Do the work. Claude implements the task and checks it (episode 4: ask for the test run).
Have it checked independently:
```
@agent-reviewer Review every change from this session (git diff). Critical issues first, with file and line. Do not change anything.
```
The blueprint's reviewer has no Edit or Write tool and is instructed not to change anything. You fix what it finds in the main session.
Save the state: paste the continuation prompt from examples/prompt-continuation.md. Claude updates current.md, lessons.md and todo.md and hands you a short prompt.
Copy, clear, continue: copy the prompt, then:
```
/clear
```
Paste the prompt and press Enter. /clear costs nothing; /compact, by contrast, summarises the history and is itself a large request (source).

Example to copy

The full continuation prompt is in the section "Your step", with an example result in examples/prompt-continuation.md.

Before the next episode

There is no next episode. End your next session with the continuation prompt and /clear, and start the new one with what Claude handed you.

If something goes wrong

You're correcting Claude for the third time on the same problem: the context is full of failed attempts. /clear and start again with a sharper prompt that includes what you learned. Source: https://code.claude.com/docs/en/best-practices
After /clear you're missing something from the old conversation: /resume brings it back. Tip: give it a name with /rename first so it's easier to find. Source: https://code.claude.com/docs/en/commands
Claude researches itself instead of using the subagent: naming it is a suggestion. The @-mention (@agent-researcher) makes sure exactly that agent runs. Source: https://code.claude.com/docs/en/sub-agents#invoke-subagents-explicitly
In a very long session Claude forgets early instructions: automatic summarising can drop details from the start. Permanent rules belong in CLAUDE.md, not in the chat. Source: https://code.claude.com/docs/en/how-claude-code-works

Free by email

With every new video you get the full guide with the ready-to-paste AI brief. Once you confirm your address, the complete Blueprint package is the first thing we send you.

Free. Unsubscribe any time.

Rather have it built for your company?

AILoopwise builds AI systems that run in your business, and keeps them running.

Book a Call

Back to the overview