slowlp
← Blog
Lesson 2026.06.30 · 7 min read

[Blog 3] Jumping from Seed Straight to Draft Kept Missing the Point

Why a two-stage seed→draft pipeline drifts off-target, and how a brief checkpoint fixed it

Lesson

When I first built the blog pipeline, it had two stages. Seed, then draft.

You register a topic as a seed file, and the AI writes a draft right away. Simple and fast, I thought. It was fast, alright. But the output kept missing what I had in mind.

What Was Off

Every time I opened a draft the AI wrote, the writing was clearly there. Topic matched, sentences were clean. But as I read, something felt off. The points I wanted to emphasize came out as generic explanations of the topic, or the part I thought mattered most was buried somewhere in the middle of a paragraph.

“Going straight from seed to draft — the AI’s just not in sync with what I’m thinking.”

This wasn’t a one-off problem.

The Real Issue: Seed Had No Direction

Inside the seed file were the topic and paths to connected raw files. A few scattered source notes too. But there was no angle for the post.

Without an angle, the AI writes the most general take on the topic. If the seed file says “blog pipeline,” the AI writes what a blog pipeline is and how to set one up. What I wanted to write was why a brief was necessary.

So the problem wasn’t that the AI wrote badly. The seed file just didn’t carry any direction. The AI wrote the most reasonable draft from the information it had, and I was frustrated that the output was different — without ever actually telling the AI what angle I had in mind.

I Inserted a Brief Stage

I added a step to the pipeline. Between seed and draft: brief.

The brief’s job is to lock in direction. It nails down the angle, the core message, and the outline. And there’s one condition: a human has to approve it before anything moves to draft.

Here’s how it actually flows. If there are source notes, the AI reads them and proposes a brief right away. If not, it puts forward three angle options. A is why-focused, B is how-focused, C is centered on the lesson or the twist — genuinely different perspectives. I pick one or tweak it to confirm the direction, and only then does it move to the draft stage.

Human approval is the gate. The AI doesn’t generate a brief and run straight to draft on its own. If I don’t sign off on the angle, it stops.

The Pipeline Started Looking Like Software

Once I plugged in the brief, the pipeline started resembling a software state machine.

seed → brief → draft → published. Each state has entry conditions. Brief requires an angle, a message, and an outline — plus human sign-off. Draft can only be written once there’s a brief. Only drafts that a human has reviewed can go to published.

The file structure is simple too. Seed, brief, and draft all live in the same file. One .md file inside export/slowloop/content-drafts/ — only the status field changes. As the state advances, fields accumulate in the frontmatter. At the brief stage, angle, message, and outline get added. No proliferating files, no confusion about which version is current.

I didn’t set out to “design this like software.” I just kept fixing problems as they came up, and it ended up this way.

How It Works Now

If a draft request comes in without a brief approval, the skill rejects it. That’s intentional.

The /blog-brief skill owns the brief stage. Run it, and it reads the seed file, then either proposes a brief based on the source notes or presents three angle options. If I say OK, the seed file’s status changes to brief and the angle, message, and outline fields get added. After that, /blog-draft becomes available.

There’s also a rule baked in: the angle can’t be built from content that isn’t in the raw files. If you pick an angle without backing, you’ll hit a wall when writing the draft. The skill has a clause requiring that the angle come from material that actually exists in the sources.

With this structure in place, when I open a draft the AI wrote, it mostly lands close to where I wanted it. Not a perfect match every time — but it almost never gets the direction flat-out wrong anymore.


In the end, it comes down to this: if you hand the AI a topic without an angle, you won’t get what you had in mind. You have to first figure out what you actually want to say — then hand it to the AI. A brief is the mechanism that forces that clarity.

From the project
Slowloop Lab LIVE
Not fast and gone — the brand hub of someone who builds, records, and builds again.
View →
COMMENTS