technical-writinglisted
Install: claude install-skill xsefirosus/sefi-agents
# Technical Writing
Craft skill backing the technical-writer. The audience is outside the team; the law is
that docs describe the tree as it is, verified.
User instructions always override this skill.
All factual output follows the anti-hallucination skill: every command, path, flag, and
number is verified against the repo or an executed output -- or marked UNKNOWN, never
guessed.
## Rule block
1. Audience first: the first lines name who this is for and what they will accomplish.
One doc, one audience -- an agent-targeted doc (Install.md) and a human-targeted doc
(README) do not merge.
2. Quickstart before theory: the shortest working path (copy-pasteable) appears before
any architecture explanation. A reader who leaves after 30 seconds should leave with
working commands.
3. Verify-before-cite: run or open everything you document. A drifted command in a
README is a bug with the same severity as a broken test.
4. Honest claims only: no invented benchmarks, no superlatives without a measurement, no
documenting features that are not in the tree. Future work is labeled as such in a
roadmap section or omitted.
5. Numbers carry sources: a figure is followed by where it comes from (a named report, a
command output). Unlabeled estimates are cut.
6. One idea per sentence; concrete over abstract; second person for instructions ("run
X"), not passive voice ("X should be run").
7. House constraints: plain ASCII (the unicode gate scans docs -- no emoji, no em