Content Creation for Website Agent

What this agent does

Drafts website content — glossary rebuilds, blog posts, capability-primer copy, template content passes — through the AI content workflow, so the content calendar ships without an external editorial agency behind it. A content lead stops writing first drafts and starts approving finished ones: the agent retrieves the brief and the source material, drafts to the house style, self-checks, and lands the piece as a CMS draft with its evidence attached. It is the succession plan for GrowthX (contract ends Nov), stated as software.

Why this agent

• The successor is committed, not aspirational — the Agentic Discovery plan says it directly: creation is "GrowthX today >> our own agent tomorrow." GrowthX proved the model for 8 months and drives ~47% of redis.io's LLM citations; when the contract ends in November, that production has to continue without the retainer.
• The volume is already scoped — the glossary program alone carries a keep / fix / retire verdict for every page plus an expansion list; the blog relaunch needs named-series velocity across three surfaces; the website build needs ~103 pages of content passes. That is more than the post-GrowthX human team can draft by hand.
• The economics are the operating model — every role runs agent-augmented; this is the flagship proof. Verto's AI engineer builds it inside the bank of hours (SOW §3.7), on the Redis-owned platform, with the hours-returned ledger tracking the payback.
• It feeds the traffic recovery — the content it drafts is the AEO counter-play: citable glossary depth, expert-led posts, new-query coverage. Citation share on Profound is how the loop closes.

Trigger & inputs

• Trigger — the content calendar (weekly plan, daily drafting queue) plus ad-hoc intake: a glossary verdict flips to *fix*, a new-query target lands from SEO/AEO, or the website build queue needs a template content pass.
• Reads — the website content module (the single copy source: nav, spine, products, proof points), the Redis voice rules and brand guide, the glossary verdict list, Ahrefs/Profound target lists, past published pieces as style examples, and engineer-validated snippets where code is involved.
• Access — individual account + Okta per the contractor security protocol; CMS draft-only permissions (publish stays human).

What it produces

• CMS drafts — publish-ready pieces in Sanity, on the right template, with metadata, schema, and internal links proposed.
• Evidence bundle per draft — sources used, claims that need verification, snippets flagged for engineer validation. Nothing asserts unbacked numbers.
• A work log — every draft, decision, and rejection recorded so the owner audits the log, not the memory.

How it works

• Pull the brief: target query/term, template, audience, and the verdict or calendar entry that triggered it.
• Retrieve context: content module copy, style rules, source material, related published pieces, validated snippets.
• Draft on the template — sentence case, verb-first CTAs, house voice — with schema and internal links included, not bolted on.
• Self-check against the quality bar: claims sourced, no invented figures, no naming violations, CTA routes to Get Started.
• Route to the human gate: CMS draft + evidence bundle to the approver; revisions loop back with the diff.

Guardrails & human-in-the-loop

• Approval gate — a human approves every publish; brand review rides the existing nine-step brief-to-production pipeline. This gate does not come off at any autonomy level currently planned.
• Never-do list — never publishes directly; never invents figures, benchmarks, or customer claims; never touches product naming (answer engines repeat whatever we publish); never edits live pages.
• Thresholds — drafting volume capped to what the approval capacity can review; queue past the cap pauses and pages the owner rather than stockpiling unreviewed drafts.
• Escalation — missing sources or a claim it can't back → stops and asks, never guesses. Cost controls and governance arrive program-wide in V3; until then spend and output quality are watched manually.

Success metrics

• North star — published pieces per week through the workflow at quality. Baseline: GrowthX-era throughput (TBD from the calendar). Target: match by GrowthX exit (Nov), exceed after.
• Quality bar — approval rate and edit distance per draft, judged by the approving lead weekly; brand-review pass rate.
• Guardrail metric — zero unreviewed publishes; verification flags caught before publish, not after.
• Downstream — citation share on Profound against the Phase 03 baseline; the agent drafts the content, the AEO program owns the outcome.
• Hours returned — drafting hours per week returned to the content leads, on the program's hours-returned ledger.

Build plan

• V0 · Shadow mode (Sep) — drafts run alongside live GrowthX production on the same briefs; output compared side by side. Exit: quality parity on a agreed sample, judged by the content leads.
• V1 · Assisted (Oct) — the agent drafts the real calendar; humans approve every piece; GrowthX still in place as the safety net. Exit: approval rate + edit distance targets hit at calendar volume.
• V2 · The succession (Nov) — GrowthX exits; the agent + the Digital Media & Content Specialist (starts Nov 1) run the workflow as the standing operation. A pilot of agent-generated content in one live campaign rides the program's V3 wave. Stays assisted by design — the publish gate is permanent.

Dependencies & risks

• Platform first — rides the V1 agent platform (repo, pods, platform decisions) shipping the same month; slippage there slips this.
• Daily owner unassigned — the H2 rule says every agent is owned like a product; assign before V1 or the rule breaks on the flagship. [owner to complete]
• Approval capacity is the real throttle — the agent can out-draft the review bench; the cap protects quality but sets the ceiling until the specialist starts Nov 1.
• Product naming gate — naming separation lands via PMM/comms before launch-window content ships; the agent inherits whatever words the platform narrative locks.