Ari’s Space*   est. 2026
Skip to content
ari*s space[ a. ]
← Librarydiscipline and process / 2026-06-21

Plan Like You Mean It

The plan is the work. How a vague idea becomes shipped software through a brainstorm, a spec, a phased plan and a typed brief, so each phase fits one session and the output is drastically better.

6 parts / Hover to explore. Click to read.

The plan is the work. How a vague idea becomes shipped software through a brainstorm, a spec, a phased plan and a typed brief, so each phase fits one session and the output is drastically better.

Idea -> understood design
                  |
            written spec
                  |
             phased plan
                  |
             phase brief
                  |
             build + check
Each document carries a different decision into the next stage. Acceptance conditions need observable checks, and a complete brief still needs review.

Before you ask for another version

You've asked for a page. The page is wrong. You explain it again, get another version, and now you're trying to remember which bits of the first one you wanted to keep.

You can spend an afternoon doing that. The conversation gets longer while the result gets less clear.

Before you change the prompt, write down what would let you accept the work. Who is it for? What should they be able to do? What must work, and what would make you send it back?

That is the start of the plan. It gives the model something better to follow than your growing frustration.

A useful definition of done tells you how to judge the next version.

Decide what you are building

I start with a conversation about the result. No implementation yet. We need to agree what we're making, what must be true when it works, and what we're leaving out.

This takes longer than typing “build me a settings page”. It also gives us a way to notice when the settings page is wrong.

You haven't lost that thinking time. You've moved it to the point where changing your mind is cheap.

Give each document a job

I use four steps. Each one leaves something the next step can use.

Brainstorm the idea. Ask what we're trying to make true and what could break it. I use /brainstorm, working through one question at a time. The aim is an understood design, not a long list of possible features.

Write the spec. Put the agreed design in a file. State the objective in a sentence you could prove wrong. Add the constraints and the definition of done. Each acceptance condition needs an action that can check it. If you can't describe that check, there's still a decision to make.

Make a phased plan. Put the work in order. Keep each phase small enough to finish in one session with room for verification and a handoff. The plan should show where one piece of work ends and the next begins.

Write a brief for each phase. Give the executor the information it needs to carry out that part without reconstructing the whole conversation.

A two-panel comparison: a crumpled one-line prompt on the left, a crisp structured brief on the right with the output sharpening
A vague prompt versus a typed brief. Same model, very different output.

Make the brief specific enough to use

A typed brief is a small product requirements document. It states the goal, constraints, acceptance criteria and phases. “Typed” means that those fields have a defined place, so an important decision is harder to forget.

For a settings page, the brief might say which settings exist, who can change them, where changes are saved, and what should happen when saving fails. That's a hypothetical example, but you can see the difference it makes. The model now has decisions to implement and behaviour to check.

Filling in those fields also makes you notice what you don't understand. You can't write a useful acceptance condition for a feature you've only half decided on.

Try the toggles below. Each one adds a missing part of the brief. The illustration shows how the direction becomes clearer; it isn't a promise that a complete brief makes every model response correct.

Prompt: make me a website

Upgrade to a typed brief:
Output preview
  • Generic boilerplate. You will be re-prompting for hours.

You'll still review the work. You now have something specific to review it against.

Leave room inside the session

A model's context window is the material it can work with in that conversation. It is finite. A project can get difficult to follow before you've filled it.

I notice this when a session starts forgetting decisions, repeating investigations or contradicting something we'd already settled. Adding more instructions to the same thread doesn't necessarily fix that.

Size the work so a phase fits comfortably, including the checks at the end. When it doesn't fit, split it. A half-finished phase needs a written account of what's working and what's left before you open a fresh session.

Try loading the whole project and then one phase in the gauge.

Memory

Load something into the session.

A row of session windows each holding a single phase floppy that fits, beside one window overstuffed and bursting
One phase per session fits. The whole project at once does not.

The plan sets the boundaries. The brief lets the next session work within them. You can build a large project without expecting one conversation to carry every detail at once.

Use the plan to find your own gaps

The useful surprise is how much you learn while writing the brief.

By the time I've turned an idea into a spec and a set of phases, I've usually found the contradiction I was ignoring, a missing decision, or something I was waving my hands at. Those are much easier to fix in a paragraph than halfway through an implementation.

The model still has work to do. So do I. But we have a shared way to tell whether that work is taking us towards the result.

For your next build, start with three things: the result, the constraints, and the check that proves it works. Then turn that into a phase small enough to finish. You can improve the prompt after you've given it a clear job.

[ comments ]

guests welcome. members show first in the list

no comments yet. start the thread_

1 / 8