sirdocalot · measured 20 August 2026 · Opus 5

What handing over costs

Comparing generating a brief with Sir DocAlot against generating an artifact with Claude Code.

The experiment

A Claude Code session had been working in this repository for a few hours — reading files, running tests, changing code. The work was finished and needed to go to a person.

We forked that session twice. Both forks inherited the same 219,172 tokens of context, so the only thing separating them was the instruction, and the instructions differed by a single sentence.

Hand the current state of this benchmark work to the operator, so they can read it at their own pace rather than out of this transcript — what was measured, what was found, what is still open, and what needs their decision.

Fork AUse sirdocalot: create a brief for them.

Fork BProduce it as a published Claude artifact they can open at a link.

When you are done, reply with only the URL and nothing else.

Three pairs ran that way. Every fork published something a person can open, and neither was told how much effort to spend.

What the brief saved

Fork A measured against fork B, median of the three pairs. Every figure is the brief costing less.

58%

Fewer tokens written

Output tokens, thinking excluded — 2.4× less. Range across pairs: 55% to 72%.

72%

Less context added

What the handover left behind in the session that ran it — 3.5× less.

65%

Less time waiting

Wall clock, from the instruction to a published link — 2.9× less.

What each fork did

Fork A — brief

1 tool call

Bash

Tokens written: 3,497
Context added: 5,138
Turns: 4
Wall clock: 61s

Fork B — artifact

3 tool calls

Skill→ Write→ Artifact

Tokens written: 8,262
Context added: 17,307
Turns: 7
Wall clock: 161s

Pair one. Fork B loaded a design skill without being asked to; nothing in either prompt mentioned design.

All three pairs

	Pair 1	Pair 2	Pair 3
Tokens written
Brief	3,497	3,194	4,413
Artifact	8,262	11,534	9,884
Ratio	2.36×	3.61×	2.24×
Context added
Brief	5,138	4,760	5,675
Artifact	17,307	24,113	19,932
Ratio	3.37×	5.07×	3.51×
Wall clock
Brief	61s	59s	66s
Artifact	161s	232s	189s
Ratio	2.63×	3.95×	2.88×
Tool calls
Brief	1	1	2
Artifact	3	5	4
Turns
Brief	4	4	6
Artifact	7	12	10

Pair 1 · briefsirdocalot.vteng.io/r/qsf28b2enmpn

Pair 2 · briefsirdocalot.vteng.io/r/2ttyff4t834b

Pair 3 · briefsirdocalot.vteng.io/r/fdj8z6q534r2

Pair 3 · artifactclaude.ai/code/artifact/b2020591…

Two artifact links are missing for the same reason: publishing an artifact from a session updates that session's artifact rather than making a second one. Pair 2's fork republished over pair 1's, and this page was later published over the same one. The numbers come from the transcripts, not the artifacts, so they are unaffected.

What this does not claim

Not that briefs look better. The published artifacts pick their own typography and lay out for their own content. A brief never will — every brief comes out looking like every other brief. That is the trade, and for something read once it is usually the right one.
Not a design comparison. Nothing here competes on bespoke design that draws its elements from the content it presents. A fixed vocabulary of widgets cannot do that and is not trying to.
Three pairs, one model, one kind of document. Pairs two and three inherited a context that already named pair one's output. Pair one is the only fully naive pair, and it is also the narrowest result, at 2.36×.
Not free. A brief costs a round trip to a service that has to be running, and a vocabulary in context. An artifact costs neither.