Context windows, memory, and Projects (beginner map)

Goal: Stop losing the plot mid-chat: know what the model can “see,” and when to start a Project instead of a forever thread.

Best for: Anyone whose chat got messy after 20 messages.

Long chats forget early rules. Put standing instructions up top—or start fresh.

Why beginners skip this

Believing the model “remembers everything” loses constraints mid-thread. Put must-keep rules in standing notes or Projects, not only turn one. Mid-week cost shows up as rework, not as a missing bookmark.

Step 1 — Three boxes (remember these names)

• Context window: the text the model can see in this chat right now (your messages + its replies + attached files). When the chat is huge, older parts may drop out — like a notepad that only keeps the last pages.
• Memory (product feature): optional facts the app saves across chats (e.g. “I prefer short emails”). Turn it off or edit it in settings if it feels wrong.
• Project: a folder-like workspace with its own files and standing instructions. Use it for ongoing work (a client, a course, a product launch).

You do not need to memorize token numbers. If answers start ignoring early instructions, the window is full — start a new chat or a Project and paste a short “handoff” summary.

Step 2 — What to click

• ChatGPT: look for Projects in the sidebar → New project → add instructions + files.
• Claude: look for Projects → create one → upload knowledge + set project instructions.
• Memory: open Settings and search for Memory — review what it stored; delete anything sensitive.

Step 3 — Write a handoff summary (copy this)

Do this step with a real artifact open—not from memory.

Handoff for a new chat/Project:
- Goal:
- Audience:
- Decisions already made:
- Constraints (tone, length, must-not):
- Files that matter:
- Next ask:

Handoff for a new chat/Project:
- Goal: Q4 onboarding email for new hires
- Audience: first-week employees, non-technical
- Decisions already made: send Monday 9am; from People Ops
- Constraints: ≤180 words; no slang; no benefits numbers (link to wiki instead)
- Files that matter: onboarding-outline.md (attached)
- Next ask: Draft v1 from the outline only

Step 4 — When to keep the same chat vs start fresh

My chat is getting confused mid-project. Propose when to start a fresh thread vs pin a summary. Context: 40-message thread about a resume rewrite; early facts were better.

Fresh thread when confused

Pin a summary; start clean. Long confused threads invent lore.

Thread is 40 turns of meeting chaos.
Build ≤15-line pack: goal, constraints, decisions, open questions.
Rerun the ask in a fresh chat with only the pack.

Red flag

Giant pastes as “memory”; critical rules only in turn one—move them to standing/Project storage.

Save this card

Card: Pack the decisions; do not paste the novel—prove it on a real artifact, or skip the badge. Keep this beside the tool; read before you start.

1. What I shipped: ___ 2. What I refused: ___ 3. What I will never-do again: ___ (date: Asia/Shanghai)

• [ ] Tools/writes paused
• [ ] Refuse-line dated
• [ ] Artifact path saved
• [ ] Next link chosen (or stop)

Before you start

Sit with three real asks from this week. You will label trust/check/stop for “Stop losing the plot mid-chat: know what”—bring the asks, not slogans. Guide: foundations-context-windows-and-memory. If you only have a blank practice box, stop and fetch the real job first.

• Goal in one line: ___
• Allowed tools: ___
• Forbidden: ___
• Human approve needed for: ___

How to use this week

This week: on one long chat, note what the model can still “see”; restart or summarize before the plot drifts. Date the restart on a live artifact—not a blank demo. Write one refuse-line and date the note (Asia/Shanghai). If you only skimmed foundations-context-windows-and-memory, put 25 minutes on the calendar and redo once.

Pitfalls

Temptation: skip the habit for foundations-context-windows-and-memory to “save five minutes.”
Job still: Stop losing the plot mid-chat: know what
Do the smallest correct pass: Done visible · one checkable artifact · Unknowns labeled [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: the artifact + the refuse-line you said aloud.

Practice now

Live card: apply “Context windows, memory, and Projects (beginner map)” to a mistake from this week; write the fix habit (Asia/Shanghai).

Today: move one must-keep rule from chat memory into a Project/standing note and prove it still applies in a new thread (Asia/Shanghai).

Today card (copy) — foundations-context-windows-and-memory: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

• [ ] Artifact path written
• [ ] Refuse example would say "yes"
• [ ] Dated Asia/Shanghai

One sentence takeaway

Pack the decisions; do not paste the novel—prove it on a real artifact, or skip the badge

• You can point to a dated refuse-line
• You have one artifact path from tonight
• You can teach the red flag in one sentence

Next

System vs user · ChatGPT Projects · Claude Projects · What LLMs can/cannot · When to switch models

Pair with today: open the first Next link (foundations-system-vs-user-prompts) only after you dated a refuse-line for this guide (Asia/Shanghai).

上下文、记忆和 Projects（新手地图）

目标： 别聊着聊着丢了线索：弄清模型「看得见」什么，以及什么时候该建 Project，而不是无限续同一个对话。

适合： 聊过二十多轮、窗口已经一团乱的人。

长对话会忘早先规则。常驻说明置顶——或开新聊天。

为什么新手跳过

整段历史粘贴显得认真。上下文窗口会悄悄丢掉中间。短包+写明「要记住什么」，胜过长篇小说。

步骤 1——先记住三个「盒子」

• 上下文窗口（context window）：这一次对话里模型此刻能看到的文字（你的话 + 它的回复 + 附件）。聊太长时，早期内容可能被挤掉——像只能留下最后几页的笔记本。
• 记忆（Memory，产品功能）：可选，把一些偏好跨对话记住（如「我喜欢短邮件」）。不对劲就去设置里关掉或改掉。
• Project：像文件夹的工作区，有自己的文件和固定说明。适合长期事（某个客户、一门课、一次产品上线）。

不用去背「多少 token」。如果回答开始无视你早就说过的规则，多半是窗口满了——新开对话或建 Project，并粘贴一小段「交接摘要」。

步骤 2——点哪里

• ChatGPT：左侧栏找 Projects → New project → 加说明和文件。
• Claude：找 Projects → 新建 → 上传资料 + 写项目说明。
• Memory：打开 Settings，搜索 Memory——看看存了什么；敏感的删掉。

步骤 3——写一段交接摘要（可复制）

用真实材料做这一步——不要凭记忆。

给新对话 / Project 的交接：
- 目标：
- 读者：
- 已做决定：
- 限制（语气、长度、禁止事项）：
- 关键文件：
- 下一句要模型做什么：

给新对话 / Project 的交接：
- 目标：Q4 新员工入职邮件
- 读者：入职第一周、非技术同学
- 已做决定：周一上午 9 点发；发件人 People Ops
- 限制：不超过 180 词；不说黑话；福利数字不写（改链到 wiki）
- 关键文件：onboarding-outline.md（已附件）
- 下一句要模型做什么：只根据提纲写第 1 稿

步骤 4——何时续聊，何时重开

聊天做到一半开始乱。建议何时新开对话 vs 钉住摘要。背景：改简历的 40 条消息线程；早期事实更好。

混乱时新开对话

钉住摘要；干净开始。又长又乱的线程会编造传说。

红旗

超长粘贴当记忆；关键约束只存在于第 1 轮。模型忘了却怪它——先查你放哪。卡上写不清就不继续——先补检查项再跑。

保存这张卡

卡面：窗口 vs 记忆 vs 项目。丢约束时看卡；关键规则迁持久处；缺项就停，不硬开。

1. 我交出了：___ 2. 我拒绝了：___ 3. 我永不再做：___（日期：Asia/Shanghai）

• [ ] 工具/写入已暂停
• [ ] 拒做句已标注日期
• [ ] 成品路径已存
• [ ] 已选下一步（或停）

开始前

打开你真正会用的账号/项目。准备好与「上下文、记忆和 Projects（新手地图）」相关的真实材料（不要空白练习框）。

默认姿态：在人确认发送/写入之前只出草稿。授予新工具前，先写一条带日期的拒做句（Asia/Shanghai）。

• 一句话目标：___
• 允许的工具：___
• 禁止：___
• 需要人批准：___

本周怎么用

本周把三条必须遵守的约束写入项目/站立说明，而不是只靠聊天记忆。把过/不过写进笔记并注明日期（Asia/Shanghai）。

周中代价检查：若明天丢掉这个习惯，24 小时内会坏在哪？写一行（Asia/Shanghai）。

常见坑

任务：Stop losing the plot mid-chat: know
场景要点：三张真实请求便签在手；练习编号 3/4。
边界：仅草稿；未知 [?]；禁止悄悄发送；禁止编造指标。
返回：可检查产出 + 过/不过理由一行。

任务：Stop losing the plot mid-chat: know
场景要点：三张真实请求便签在手；练习编号 4/4。
边界：仅草稿；未知 [?]；禁止悄悄发送；禁止编造指标。
返回：可检查产出 + 过/不过理由一行。