Your AI Coach

Your AI Coach

The admin chat is your private workspace for improving your agent. It's not a text editor — it's an AI coach that reads your soul, researches the web, analyzes visitor behavior, and proposes specific changes for your review.

The Proposal Workflow

The admin coach never edits your soul directly. Every change goes through a proposal workflow:

1. Read — coach reads your current soul to understand what's there
2. Propose — coach generates a specific change (edit or rewrite)
3. Review — you see the proposal in a ProposalCard with the exact changes highlighted
4. Accept or Decline — you decide. Nothing changes until you approve.

This is intentional. Your soul is too important for unreviewed automated changes.

Surgical Edits vs Full Rewrites

Surgical edit: Changes a specific section while leaving everything else untouched.

• "Add a rule about handling pricing questions"
• "Update my job title from Senior to Staff Engineer"
• "Add an edge case for when visitors ask about competitors"

Full rewrite: Regenerates the entire soul from scratch.

• "Rewrite my soul with a more casual tone"
• "Start over — I want a completely different approach"
• "Regenerate my soul based on my updated LinkedIn"

The coach decides which approach to use based on your request. Small, targeted requests get surgical edits. Broad requests get rewrites. You can also explicitly ask: "Do a surgical edit, not a rewrite."

Research Workflow

The admin coach can search the web to inform soul improvements:

Example conversation:

You: "I want my agent to know about my recent conference talk." Coach: searches for your conference talk, reads the page, proposes adding key talking points to the Knowledge chapter

Example conversation:

You: "Update my product knowledge based on our current website." Coach: fetches your website, extracts current product information, proposes replacing outdated Knowledge section

The coach uses the same search pipeline as your agent (Jina + Brave parallel search, deep URL fetching). It can read any public URL you point it to.

Visitor Insights

The admin coach can pull analytics about your agent's visitors to make data-driven suggestions:

You: "What are visitors asking about most?" Coach: queries visitor analytics, finds top topics, suggests adding Knowledge or adjusting response routing

This turns raw conversation data into actionable soul improvements. If 40% of visitors ask about pricing and your soul doesn't have a pricing section, the coach will suggest adding one.

Version History

Every soul change creates a new version. You can:

• View the last 20 versions — see when each version was created
• Read any version — view the full soul prompt for any historical version
• Activate any version — restore a previous version without losing anything (non-destructive)

Version history is accessed through the admin panel's version history sidebar. Each entry shows the version number, creation timestamp, and a preview.

How activation works: When you activate a historical version, it doesn't delete newer versions. It sets activeSoulVersion on your agent to point to that specific version. Your agent will use that version's prompt until you create a new version or activate a different one.

This means you can experiment freely — if a soul change doesn't work out, just activate the previous version. You'll never lose a working soul.

Coach Memory

The admin coach remembers your preferences across sessions:

• Editing style: Does she prefer surgical edits or rewrites?
• Accepted/declined patterns: What kinds of proposals do you accept? What do you decline?
• Explicit preferences: "Never suggest emoji" → remembered forever
• Open threads: What improvements were discussed but not yet applied?

Coach memory reflects every 5 conversations, compressing raw interaction logs into a structured "Creator Profile." This means the coach gets better at anticipating your preferences over time.

Saving explicit facts: If you tell the coach something important — "I never want bullet points in my agent's responses" — it saves that as an explicit fact in coaching memory. These facts persist permanently and are referenced in every future session.

The Remember Command

You can explicitly tell the coach to remember something:

"Remember: I always prefer short, punchy responses over detailed ones." "Remember: Never suggest adding emoji to my soul." "Remember: My agent should never discuss competitor products."

These are saved as permanent facts in your coaching memory and influence all future proposals.

Effective Coaching Conversations

Be specific about what you want:

• "Make my responses shorter" → coach has to guess what "shorter" means
• "Limit casual responses to 1 sentence, technical to 3" → coach knows exactly what to do

Reference specific problems:

• "My agent doesn't sound right" → too vague
• "My agent says 'Great question!' at the start of every response — fix that" → actionable

Ask for research:

• "What would make a good fact-checker soul?" → coach searches for best practices and proposes changes
• "Find my latest blog post and add it to my knowledge" → coach fetches, reads, proposes

Use version history:

• "Compare my current soul to version 3" → helps understand what changed
• "My agent was better last week — show me what changed" → coach can diff versions

Common Admin Chat Patterns

1. Initial soul review: "Read my soul and suggest 3 improvements"
2. Adding knowledge: "I just published a new blog post at [URL] — update my knowledge"
3. Voice tuning: "My agent sounds too formal — propose voice changes"
4. Edge case discovery: "What questions is my agent struggling with?" (uses visitor insights)
5. Competitive research: "Search for how other AI agents handle [topic] and suggest improvements"
6. Quick fix: "Add 'No emoji' to my voice rules"