Website Review Runbook (Multi-Agent)

Version: 1.0 Scope: all published pages by default.

Inputs to provide per run

Create a context packet using work/website-review/runbook/context-packet-template.md.

Required context packet fields (must be filled by the orchestrator):

• Review scope: all pages (or explicitly constrained run scope)
• Pages: explicit page list for this run
• Code style guide: context/code-style-guide.md
• Tone/voice style guide: context/tone-voice-style-guide.md
• What changed: copy/UI/structure changes since last review
• Audience assumption: executive recruiters and hiring managers
• Success definition: recruiter/hiring manager can understand value quickly and take a low-friction next step (contact) with credible proof
• Known constraints: static HTML/CSS/JS; follow both guides for recommendations

Fixed run set (ALWAYS run; do not omit)

Personas (always all 6):

• ceo-exec
• chief-of-staff-recruiter
• technical-ceo
• startup-exec-cofounder
• people-ops-hr
• culture-values

Disciplines (always all 4):

• ux-content
• ui-design-systems
• code-semantics
• seo-accessibility

Workflow phases

Phase 0: Orchestrator setup

1. Confirm the page list is exactly the Pages: from the context packet.
   • Default page set (if not constrained): index.html, blog/index.html, projects/index.html, and projects/*/index.html.
2. Load the per-agent definition files from work/website-review/agents/:
   • Each agent must receive ONLY:
     • the filled context packet
     • the page(s) to review (default all pages)
     • context/code-style-guide.md
     • context/tone-voice-style-guide.md
     • its own definition file from work/website-review/agents/
3. Tell agents:
   • Do not reference other agents.
   • Do not add “generic advice.” Every issue must be tied to a specific section/phrase/element on the page.

Phase 1: Persona Scan

Run all 6 persona agents in parallel.

For each persona agent:

• The agent must follow its own output schema exactly (see its definition file).

Phase 2: Discipline Deep-Dive

Run all 4 discipline specialists in parallel.

For each discipline specialist:

• The agent must follow its own output schema exactly (see its definition file).

Phase 3: Synthesizer merge + dedupe + prioritize (keep it thorough)

Synthesizer must:

1. Dedup only exact duplicates and same-root-cause repeats.
   • If agents report different root causes (even if they feel similar), keep them separate.
2. Preserve thoroughness: do not compress by deleting evidence; shorten only by deduping exact duplicates.
3. Output a single final report with:
   • Executive Summary: 3-6 bullets
   • Categorized Findings: grouped into:
     • UX/Content
     • UI/Design
     • Code/Semantics
     • SEO/Accessibility
     • Persona/Resonance
   • Prioritized Action Plan: ordered buckets:
     • Must (ordered)
     • Should (ordered)
     • Nice (ordered)
   • For every Must item include:
     • Why now: 1 sentence
     • Implementation notes: 1-3 bullets