Huzaifa Tofeeq

[Huzaifa Tofeeq @huz_tfq]

When you have to kill the session with all the good context because Claude lost the plot.

[Huzaifa Tofeeq @huz_tfq]

@betomoedano nice one

[Huzaifa Tofeeq @huz_tfq]

Developers using AI daily: what’s the one thing you still do manually every single session?

For me it’s explaining what I was doing before I opened the chat.

[Huzaifa Tofeeq @huz_tfq]

@leo_mercerr Just started on X, hoping for the best

[Huzaifa Tofeeq @huz_tfq]

@Kedeodev If it runs on auto and you can focus on other things then yes

[Huzaifa Tofeeq @huz_tfq]

@BabaDev_ Vercel, handles everything

[Huzaifa Tofeeq @huz_tfq]

@kushmergedeck happy to connect

[Huzaifa Tofeeq @huz_tfq]

Software engineering has officially shifted from typing out every line to being a very stressed-out air traffic controller for LLMs.

[Huzaifa Tofeeq @huz_tfq]

Asked how people carry context between AI coding tools. Almost everyone said a notes file in the repo.

The part most people skip: what already failed.

Without it, the next model happily retries the same three fixes you ruled out an hour ago.

[Huzaifa Tofeeq @huz_tfq]

@thesimaximalist “Keep it updated as you work” is the key part. Nobody remembers to write it once the limit has already hit.

[Huzaifa Tofeeq @huz_tfq]

Question for people using more than one AI coding tool:

When Claude Code / Cursor / Codex hits a limit mid-task, how do you carry the context over to the next one?

Copy-paste the chat? A notes file? Just start over?

[Huzaifa Tofeeq @huz_tfq]

@actvox Agreed, the dead ends are noise, except for a one-line “tried X, didn’t work” so nobody retries it. I’d keep it in the repo root. Most tools look there first.

[Huzaifa Tofeeq @huz_tfq]

@SolutionsCay One session per task is a clean way to avoid the problem completely. Is the ledger something you built yourself, or an existing tool?

[Huzaifa Tofeeq @huz_tfq]

@kylekane Nice, a single source of truth. Do the agents write to it on their own, or do you update it yourself?

[Huzaifa Tofeeq @huz_tfq]

@radialbuild “Starts at attempt six” is the key part. Logging what failed matters more than logging what worked, because otherwise the next model tries the same fixes again.

[Huzaifa Tofeeq @huz_tfq]

Claude Code hit its usage limit mid-task again.
Switched to Gemini and spent 15 minutes re-explaining the bug, the files, what I’d already tried.
The model isn’t the bottleneck anymore. Context is.

[Huzaifa Tofeeq @huz_tfq]

@automater_ai Agreed. The trick is getting it written before the limit hits, not after. Do you have the agent update it every step, or only at the end?

[Huzaifa Tofeeq @huz_tfq]

@Surendar__05 32 at least, or max 48
64 is just overkill

[Huzaifa Tofeeq @huz_tfq]

@Tanju_mim @marryevan999 G

[Huzaifa Tofeeq @huz_tfq]

@marryevan999 n8n