Context Management 105: ย่อและ refresh contextContext Management 105: compress and refresh

Long AI conversations get messy.

At the start, the model sees a clear task. After many turns, it also sees old drafts, rejected ideas, corrections, tool outputs, and side discussions.

Compression is how you keep the useful parts and remove the noise.

When to compress

Compress context when:

• the chat is long
• the answer starts contradicting earlier decisions
• you are switching from research to writing
• you are switching from planning to implementation
• old options were rejected
• you need another session or teammate to continue

Compression is not just making text shorter. It is choosing what must survive.

What to keep

A good compressed context keeps:

• goal
• current state
• locked decisions
• source boundaries
• important constraints
• unresolved questions
• next action
• files or links that matter

It removes:

• duplicate discussion
• rejected options that no longer matter
• raw tool output that has already been summarized
• emotional back-and-forth
• stale assumptions

The working-state summary

Use this template:

Current working state:

Goal:

What is already decided:

Important context:

Constraints:

Open questions:

Next step:

Do not rely on:

The final line is powerful. It explicitly kills stale context.

Example:

Do not rely on:
- the earlier idea to make this a paid workshop
- the old title "AI Context Tricks"
- the first outline with only three lessons

Decision log

A decision log is a small list of choices and why they were made.

Locked decisions:
- Use "Context Management 10x" as title because it matches the free workshop naming system.
- Use six lessons because nearby free workshops use 101-106 progression.
- Keep Thai as default because the academy audience is Thai-first.

Do not write a long diary. Future you needs decisions, not every conversation turn.

Refresh the task

After compressing, restart with a clean instruction:

Use the working-state summary below as the current source of truth.
Ignore earlier rejected ideas unless they are listed here.
Continue from the next step.

Then paste the summary.

This is better than letting the model guess which old messages still matter.

Ask AI to compress itself

Use this prompt:

Compress this conversation for continuation.
Keep only:
- goal
- current state
- locked decisions
- important source boundaries
- unresolved questions
- next action
- warnings or stale assumptions to ignore

Remove duplicate discussion, rejected ideas, and raw intermediate output.
Return a concise handoff summary.

Then read the summary before trusting it. Models can accidentally drop subtle but important details.

Practice: compress a messy thread

Take a real or imagined messy chat.

Write:

Current working state:
Goal:
What is already decided:
Important context:
Constraints:
Open questions:
Next step:
Do not rely on:

Then ask:

Based only on this compressed context, what should happen next?

If the answer makes sense, your compression worked.

Next: Context Management 106 - create a context pack.