DOCS-988 - "Humanize" GitHub/Jira comments

[kimsauce]

Purpose of this pull request

A comment Claude Code posts should read as though a person wrote it. This makes that a requirement for GitHub PR comments, GitHub issue comments, and Jira ticket comments.

• .claude/skills/humanizer/SKILL.md vendors blader/humanizer v3.0.0 (MIT), copied verbatim so version bumps stay a clean overwrite. It rewrites the patterns that mark prose as machine-written, working from 25 numbered tells drawn from Wikipedia's "Signs of AI writing", and leaves code, commands, paths, and link targets untouched.
• AGENTS.md gains an Output tone section requiring the skill on the draft of every comment, in its embedded mode so only the final text comes back. It also states three rules that hold whether or not the skill runs: no em dashes, be concise, plain language. On a PR, one more rule applies: put each point in an inline comment on the line it concerns, and keep a top-level summary comment only for feedback not tied to a specific line.

The rules sit alongside the skill rather than inside it because a skill invocation is model-initiated, and so probabilistic. If the skill does not fire, the rules still apply.

Pairs with DOCS-1867 (#7099), which removes the attribution marker Claude appends to comments. Between them: the marker that labels a comment as automated, and the style that betrays it anyway.

Select the type of change

• Minor Changes - Typos, formatting, slight revisions
• Update Content - Revisions, updating sections
• New Content - New features, sections, pages, tutorials
• Site and Tools - .clabot, version updates, maintenance, dependencies, new packages for the site (Docusaurus, Gatsby, React, etc.)

Ticket (if applicable)

https://sumologic.atlassian.net/browse/DOCS-988

🤖 Generated with Claude Code

[amee-sumo]

Reviewed the vendored skill and the AGENTS.md changes:

• SKILL.md is byte-identical to the documented upstream checksum.
• The "Output tone" section is well-scoped to comments, and the em-dash rule is consistent with the existing sumo-style doc-prose convention rather than a new invention.
• The skill's process explicitly requires keeping every supported claim and never inventing sources, so this is a style pass, not a content-fabrication risk.
• Enforcement is instruction-only for now (no hook), which the ticket's own implementation notes already call out as a known follow-up once #7099's hook lands on main.

Non-blocking note: paired with #7099 (which drops the "via Claude Code" marker), this means a GitHub/Jira comment posted under a contributor's own account will carry neither an attribution marker nor a stylistic tell. That may be exactly the outcome intended from the 9/2 Tech Docs discussion, but it's worth the team double-checking that the "no way to tell AI-drafted from human-drafted after the fact" consequence was the explicit intent, not just a side effect of "make comments sound less robotic."

Approving.