Fork your conversations and rebase your prompts

..

2026-05-24

Fork your conversations and rebase your prompts

One of the skillsets that the AI phenomenon has quietly boosted is the capability of explaining yourself : putting into words what you actually want, and providing methodologies that can unambiguously verify the implementation against your expectations. Turns out that if you can’t describe your intent in a way that survives a stranger reading it cold, the agent can’t either. The good news is that this skill compounds. The bad news is that most of us discover we suck at it the first time we type a prompt and the agent produces something that sounds aligned with your intent, but is fragile and full of disliked side effects.

Your prompt is a lossy encoding of your intent

The agent’s job, stripped to its bones, is to infer your intent from very limited sources. Your intent is the real message: rich, fuzzy, packed with implicit context you take for granted. Your prompt is just its encoding , and it is almost always a lossy compression of that intent. You, as the user, are also the encoder of this message, and that means you carry the responsibility of encoding it properly to increase the chances of an accurate decoding on the other side. Nobody else is going to do this job for you, and the agent certainly can’t read your mind no matter how many parameters it has.

By default, most agents try to amplify the signal of your prompt by evaluating contextual information around it: the folder you are currently working on, the previous messages exchanged, system prompts, user-defined skills and rules. When this contextual scaffolding is solid and your prompt has high quality, the amplification works in your favor and the agent reconstructs something close to your original intent. When it isn’t, things fall apart in spectacular ways: if the surrounding context is low quality, and/or your prompt is noisy due to vagueness and contradictions, what gets amplified is noise . The output is then the canonical AI slop: confident, plausible-looking, and structurally wrong.

The upshot is that agentic engineering requires actual effort to safeguard the quality of prompts and contextual information, and this is not as easy as it might sound. There are a lot of blind spots and missing pre-requisites for an external listener hearing your idea for the first time, and you, while pitching your intent, might genuinely fail to guess all of the information that is missing on the other side. You can’t audit something you don’t know is missing . Typically, this gap is handled naturally during the conversation thread, but only on one condition: that the agent is proactively encouraged to be opinionated about your ideas and to raise questions when it feels unsure. Without that nudge, it will happily fill the void with plausible guesses, and you will only notice once the implementation is already on fire.

Ask for agent feedback

When you are encoding your message (i.e. writing your prompt) you are inevitably estimating which information the agent will need in order to decode it, but your estimate can fail in exhaustivity. So asking the agent for feedback on how confident it is in understanding your intent is essentially asking for a residual : the gap between what you transmitted and what the agent actually needed. That residual is valuable information for you, because it tells you exactly where your encoding leaks, and it is typically used in follow-up messages to clarify your intentions and patch the missing bits.

The problem of clarifying your intent with multiple prompts

This approach is very workable, but it has a cost. Your message (your intent) ends up diluted across multiple lossy encodings (multiple messages), and while effective, this is quite inefficient because it erodes a fundamental budget constraint in agentic engineering: the context window . After a long conversation thread the agent will start context-rotting, and this will probably kill the quality of the agentic contribution. Worse, encoding your intent across multiple prompts also lowers the quality of the contextual information that the agent will lean on in the future, when it tries to amplify the signal from the next messages: the past noise is now part of the surrounding context. The takeaway is that high-quality prompts generate positive externalities both for the current implementation and for the related follow-ups down the line .

Don’t reply, rewind: fork your conversation

In order to mitigate this, I started doing a stupid-simple thing that improved the quality of agentic contributions a lot. I just append the following at the end of my prompt:

Before starting the conversation, return your confidence level in the assignment understanding. If it is below 100%, tell me which clarifications you need (if any) and if you have divergent ideas (if any) be opinionated about it, otherwise start the implementation.

I noticed that the agent will...