Teach one next action per Needs-you entry

[than]

An entry in the human-blocker section is one next action, GTD-style: one physical step, in one place — one tool or site, one verb — that the human can finish in one sitting. The CLAUDE.md note and starter template now say so, in the same paragraph as the existing Next: rule:

• A Next: line that says "and" or names two places is two entries.
• Each entry carries what the human needs to act: the direct link, and the exact values to paste, so nothing has to be looked up.

The rule names the section by its own heading, so it disappears on a board with no human-blocker section rather than pointing at one that isn't there. The note is only refreshed when sidecar init is re-run in a repo, which the README now says.

Not included: a lint that nudges the agent when a Next: line has "and" or two URLs. Matching "and" would split things that are one action; two different hosts is a better signal and can follow if the note alone doesn't change the behavior.

go test ./... passes, including a test that the rule appears with the human section and vanishes without it.

🤖 Generated with Claude Code

[claude]

Review: solid, ready to merge.

This PR only changes prompt text, so none of the viewer's core behaviour is touched: no changes to watching, reload/scroll, render width, or styling.

Correctness

• The new rule is inside the if human, ok := humanSection(...) block, so it uses the same role lookup as the placement rule. It disappears when the board has no human section and follows renamed or re-emoji'd headings. The test covers both cases.
• Because entryStyleRules is shared, the starter template and the CLAUDE.md note can't drift apart.
• The README sentence is accurate: writeClaudeNote → replaceClaudeNote rewrites the text between the markers on re-init.

Nits (optional)

• Secrets: "the exact values to paste" will sometimes mean tokens or keys. The board is git-excluded, but its contents also get copied into .sidecar/previous.md and can show up in sidecar diff hook output, which puts them in the agent transcript. A short qualifier would stop agents from writing credentials there, e.g. "the exact values to paste (never secrets — say where to get them)".
• Possible conflict with the size cap: the shape list allows at most "two sentences of detail", but a block of values to paste could be longer. An agent might drop values to stay under the cap, or ignore the cap. If values should be exempt, say so in the same sentence.
• The README calls it a "🧠 entry", but users can re-emoji sections. That's fine as a reference to the default; mentioning it only because the code itself is careful not to hard-code the heading.

The new test is focused and checks the right things.