Free · copy-paste · about 1 minute

Your AI gets dumber
the longer you talk.

The fix isn't a smarter model. It's a fresh chat that already knows where you left off.

Every answer in a chat gets read against everything that came before it. The longer one chat runs, the more old back-and-forth it's wading through, and the answers go vague, repetitive, or just wrong. People blame the model. It's the chat.

The fix is boring and it works: start a fresh chat, and hand it a short brief so it picks up exactly where the old one died. This is the block I paste, every working session, every day. Fill the brackets, paste it as the first message, and you're back in business without scrolling through an hour of history.

Watch first

The two-minute version

Here's the whole thing, start to finish. Then grab the block below.

The whole thing

The brief I paste

Five fields. Fill the brackets, paste it as the first message of a new chat, and the new session starts already caught up. That's the entire move.

The fresh-start brief
You are picking up work in progress. Read this brief before doing anything.

GOAL
[What I'm trying to get done, in a sentence or two. The finish line, not the task list.]

CURRENT STATE
[Where things stand right now. What exists, what works, what's decided. Paste the real thing, the draft, the code, the plan, instead of describing it.]

WHAT CHANGED
[What happened since this started. Decisions made, direction changes, anything that turned out different than expected.]

WHAT FAILED
[What we already tried that didn't work, and why if you know it. Don't suggest these again.]

NEXT STEP
[The one thing to do first. Start here.]
DOWNLOAD THE BRIEF (.txt) · KEEP IT HANDY
The whole flow

Five steps, any time a chat goes stale

STEP 1
Chat gets long
answers go vague
STEP 2
Don't scroll up
start a fresh chat
STEP 3
Paste the brief
five fields
STEP 4
It catches up
instantly
STEP 5
Back to sharp
keep going
Not a vibe

This is measured, not a hunch

Researchers gave it a name: context rot. Test the top AI models on a long chat versus a short one and accuracy falls off a cliff. Here's how far the big models dropped once the chat hit about 32,000 tokens, a long but normal working session:

ModelShort chatLong chat (~32K)
GPT-4o99.3%69.7%
Gemini 1.5 Pro92.6%48.2%
Llama 3.3 70B97.3%42.7%
Claude 3.5 Sonnet87.6%29.8%

Accuracy on a long-context recall test, short chat vs ~32K tokens. 10 of 12 models tested fell to half their short-chat score or worse by 32K. Source: NoLiMa, Adobe Research.

Most of them started slipping way earlier than that, some by just 2,000 to 8,000 tokens. And it isn't only the older models above. Anthropic's own developer tool reports that even the newest one, Claude Opus 4.8 (a million-token window), starts slipping around 40% full and flags you to start a fresh session by about 48%. That's the exact line I draw: reset before you're 40 to 50% in.

Newer models raise the ceiling. They don't remove the floor. The fix isn't waiting for a smarter model. It's the fresh chat above.

Sources: NoLiMa (Adobe Research) · Chroma, "Context Rot" · Claude Code degradation report

Know the moment

When to reach for it

01

The answers start slipping

Vague, repetitive, or flat-out wrong. That's the rot, not a bad day for the model. Start fresh.
02

You're switching tasks in the same project

New task, new chat. Don't drag the last thing's whole history into the next one.
03

You walked away frustrated

Come back to a fresh chat, not the one you rage-quit. Paste the brief and pick up clean.

Rule of thumb: serious work gets a fresh chat early and often. Don't wait until it's obviously broken.

Do it right

Three things that make it work

01

Write it at the end of the old chat, not from memory

Before you close the dying chat, tell it: "Summarize this work as a handoff brief: goal, current state, what changed, what failed, next step." Fix anything it got wrong. It fills the block better than you will an hour later.
02

Paste real material, not descriptions

"The draft is pretty good" tells the new chat nothing. Paste the actual draft, the code, the plan. The brief points to the real thing, it doesn't describe it.
03

One brief, one goal

Two projects is two fresh chats with two briefs. Mixing them in one chat is exactly how the rot starts again.
Advanced, optional

Want it on autopilot? Make it a one-command skill

Skip this if you're just getting started. The copy-paste block above is all you need. But once you've pasted it a few times, you can turn the whole move into a single command.

If you use Claude Code (the terminal version of Claude), a skill is a saved instruction Claude runs on demand. I built one called /handoff. At the end of a session I type it, and it reads the whole chat, fills the five fields for me, saves the brief, and hands me the block to paste into the next one. No copy-paste, no writing it from memory.

01

Get the skill-creator (it builds the skill for you)

skill-creator is a free, official Claude plugin. In Claude Code, run /plugin and install skill-creator from Anthropic's marketplace, or run /plugin install skill-creator@claude-plugins-official. Plugin page: claude.com/plugins/skill-creator.
Then run /skill-creator and tell it you want a new skill called handoff. It scaffolds the files so you don't hand-write them.
02

Give it this job (paste this)

Tell skill-creator exactly what the skill should do. Here's mine, word for word:
Paste into skill-creator
Create a skill called "handoff".

When I run it, it should:

1. Read the CURRENT conversation (not memory) and write a fresh-start brief from it.

2. Fill these five fields, specific over vague:
   GOAL: the finish line in a sentence or two, not a task list.
   CURRENT STATE: what exists, what works, what's decided. Paste the small thing being iterated (a draft, a snippet) and reference big files by path.
   WHAT CHANGED: decisions made, direction shifts, anything that turned out different.
   WHAT FAILED: what we already tried that didn't work, so it's not suggested again.
   NEXT STEP: the one thing to do first.
   Then a SUGGESTED TOOLS line: which tools or skills the next session should reach for first.

3. Reference, don't duplicate: point to files and links by path. Only paste material that lives nowhere but this chat.

4. Never include passwords, keys, or anything private. Reference where the secret lives instead.

5. Save it as a dated markdown file in my notes, and print the filled block fenced so I can paste it as the first message of a new chat.
03

Two rules that make it good

Build the brief from the actual conversation, not from memory. And never put a password, key, or anything private in it. Point to where the secret lives instead.

Now closing a session is one command, and the next one starts already caught up. (No terminal? The copy-paste block above is your version. It works just as well.)

Quick answers

Before you ask

This is how I actually run every working session with AI.

I put out one of these a week. The boring, useful part of AI for people who run real businesses. No hype, no jargon.

Want your whole business wired like this?→

The brief is the beginner move. The real thing is wiring your business so AI runs parts of it. That starts with a free call, no pitch.