Specline canon 2.3.0-draft Handbook For agents ↗
Why Specline · Why now

Why now

The usual worry is that a spec is scaffolding a strong model has outgrown. It's the opposite: Specline helps more as the model gets stronger, because the thing it rations — judgment, context, and a falsifiable definition of done — is exactly what even a frontier-class model still can't supply for itself.

One-shot is a property of the pair, not the model

Frontier models advertise first-shot correctness on complex, well-specified problems. Read the qualifier. One-shot isn't the model reading your mind — it's the model, given a problem specified well enough that the answer is determined, producing it in one pass instead of ten. The bottleneck didn't vanish when models got better; it moved — out of the model's reasoning and into the quality of the specification handed to it.

One-shot is a property of (model × spec), not of the model. Specline manufactures the spec half.

A vague prompt to a frontier model doesn't one-shot — it confidently builds the wrong thing, or over-builds, because nothing told it not to. Intent stated, mechanics omitted; non-goals explicit; reserved decisions made up front; a falsifiable done. That list is the spec body.

The vendor's own guidance is Specline's boundaries

The strongest evidence the methodology is aimed correctly: a frontier model's official prompting guide, written with no knowledge of Specline, independently prescribes the same moves.

Frontier-model guidanceSpecline
"Give the reason, not only the request."Intent over description (B6)
"State the boundaries… don't add features beyond the task."The Non-goals section
"Fresh-context verifier subagents beat self-critique; verify against the spec."Agent-loopable acceptance checks + the reviewer
"Construct a memory system — one lesson per file; delete wrong notes."status.md + graduated knowledge
"Pause for the user only on a destructive action, real scope change, or input only they can give."The two human gates

When the model vendor converges on your methodology without having seen it, the methodology is tracking something real about how these models work — not repackaging an older process.

The one real risk

The same guides warn that instructions "too prescriptive" for a frontier model degrade its output. So Specline's one failure mode is violating its own rule: a mechanics-heavy spec — restating what the code will do, enumerating steps the model would choose better itself — measurably lowers a frontier model's quality. Intent over description stops being a tidiness preference and becomes a performance requirement. Specline helps exactly to the degree it stays intent-heavy and mechanics-sparse.

The verdict

Specline supplies the three things a frontier model can't self-source — reserved judgment, the right context, and a falsifiable done — and stays out of the way on mechanics, which the model now does better than any spec could dictate. The margin widens with capability. One-shot isn't a rival to this methodology; it's the same claim from the other side — and producing the "well-specified" half, cheaply and with the model co-drafting, is precisely what Specline is for.