Write one brief in three voices, and compare

The fastest way to choose a voice is to hear the same piece in several. This takes one sentence with an agent, or a handful of commands by hand, and leaves you with three drafts to read.

With an agent

Create three drafts in the "Field notes" collection titled "Why remote work is here to stay". Have Hazel, Aspen and Cedar each write one from this outline: cost savings for employers; the talent pool goes global; the productivity debate; what hybrid gets wrong. Don't publish anything. Tell me when all three are done and show me the opening paragraph of each.

The agent will call list_kernels to find the three ids (each kernel comes with a one-line description of what it writes like, so an agent can pick on that rather than on the codename), ingest_article three times, run_kernel (with each article_id) three times (the runs happen in parallel, since each locks only its own draft), poll get_kernel_run_status until each is done, then get_article to read them back.

By hand, with the CLI

letterstory kernel list                       # note the three kernel ids
letterstory kernel endpoints                  # note a kernel endpoint id

for K in <hazel-id> <aspen-id> <cedar-id>; do
  A=$(letterstory posts new --title "Why remote work is here to stay" \
        --collection <collection-id> --body "(draft)" --json | jq -r .article_id)
  letterstory kernel write <endpoint-id> --article "$A" --kernel "$K" --file brief.md
done

letterstory kernel status <kernel-job-id>     # repeat until status: done, applied: true
letterstory posts show <article-id>

brief.md is your outline:

Why remote work is here to stay

- Cost savings for employers
- Talent pool goes global
- The productivity debate
- What hybrid gets wrong

What to look at when you compare

• The opening. Does it start on the claim, or clear its throat?
• How it handles the weakest point. Every outline has one; a good voice makes it interesting rather than padding it.
• Sentence length and rhythm. Casual voices vary it; authoritative ones keep it steadier.
• The run page. Each result's run_url shows the rounds it took and how the reviewer panel voted. A voice that passes in one round on your material is a voice that fits it.

When you've chosen, set it as the organization default (KB-702) so Auto-Pen and the Planner use it without being told. If none of the three is quite right, build your own (KB-705) and put it in the same comparison.