Tip 01
Long chat into a skill.Instead of: Long chat for everything.
Your AI isn’t getting dumber. Your chat is getting longer.
A long chat works at first. That’s why we keep it. But slowly it turns into broken telephone.
Every new message lands on top of everything before it: old drafts, detours, rules you changed your mind about. Your first instructions end up buried in the middle, the part models are worst at using. And when a chat gets very long, some apps squeeze the oldest messages into a summary. If it fills up again, that summary gets squeezed too. Each round, a few details fall out.
Don’t let your chat become broken telephone:
- One chat per task.
- When a chat finally gets it right, ask: “Write the rules we settled on as a short checklist I can reuse.”
- Save that checklist. Claude calls it a skill, ChatGPT has project instructions, or a plain note works too.
- Next time: a new chat, plus the checklist, plus today’s facts. Your message arrives first-hand, not passed down the line.
- Before
- One chat for your whole job hunt. By the sixth cover letter, it can slip in a detail from the second company.
- After
- Save your CV facts and style rules once. For each job: new chat, paste the rules and that one job ad.
Worth watching and reading
- Andrej Karpathy, “How I use LLMs” Video, from about 16:30: start a new chat when you switch topic.
- Anthropic, “Effective context engineering for AI agents” Why context is a limited resource.
- Chroma, “Context Rot” Tests showing performance drops as input gets longer.
- Liu et al., “Lost in the Middle” 2023: models use the start and end of a long input best.