CLAUDE.md

Operating system for DaisyAI — a two-person startup building AI tooling for healthcare utilization review.

This repo is not a codebase. It's a business operating system stored as markdown files. There's a Next.js dashboard that reads some of these files, but the files are the source of truth.

---

DaisyAI

Clinical case management for healthcare utilization review (UR). Helps nurses evaluate medical necessity with AI-driven, criteria-based reviews.

Current mode: FDE (forward-deployed engineering) consulting with health plans, with a path to productization. First client: Premera (Blue Cross, West Coast).

Related repos:

• daisyai-ai/daisyai-app — Main product (Next.js 15, React 19, TypeScript, Neon Postgres, Clerk, Vercel AI SDK)
• daisyai-ai/daisyai-www — Marketing site

---

Stable Systems

These are established and have their own READMEs with full documentation.

Network (network/)

Relationship management. Everyone we actually know — advisors, mentors, connectors, investors, warm contacts. Each person gets a directory with profile.md and optionally calls/ for conversation history.

README: network/README.md

When someone new comes up in a transcript or conversation, create their directory here. When a call happens with someone in the network, call notes go in their calls/ folder and the interaction table in their profile gets updated.

Clients (clients/)

Active client engagements. Three-layer pipeline: raw inputs → structured sub-items (questions, scope, commercial) → synthesis overview. Each client gets a directory.

README: clients/README.md Active: Premera, McBee

Fundraising (fundraising/)

Investor outreach, materials, meeting notes, and cross-call analysis. Per-firm tracking in investors/, meeting notes in meetings/, synthesized patterns in analysis/.

README: fundraising/README.md Status: Raising $2-3M first institutional check. Multiple investor conversations active.

Outreach (outreach/)

Cold/warm outreach campaigns to health plans. Experiment-driven — each campaign has a hypothesis, message, and results. Company profiles and outreach history tracked per-target.

• campaigns/ — numbered experiments (001-health-affairs-cmo, etc.)
• companies/ — target company profiles + outreach records
• learnings.md — rolling synthesis across campaigns

---

Other Directories

These exist and are used, but the structure is more fluid.

Ops (ops/)

Operational hub. Priorities, huddle records, and personal logs.

• objectives.md — Current sprint priorities and task tracking. Check this first for context on what matters right now.
• huddles/ — Daily sync summaries. Processed output only: decisions, key context, action items. Raw transcripts stay out — the value is in the synthesis. See "Huddle Processing" in Agent Behavior.
• hot-board.md — Things that need action in the next few days
• personal/{michael,thomas}/ — Daily work logs per operator
• transcripts/processed.json — Tracks which transcripts have been processed

Content (content/)

Marketing content pipeline — drafts, published posts, ideas backlog. Organized by platform (Twitter, LinkedIn, Substack).

• pillars.md — Voice guide and 5 content pillars
• drafts/{platform}/ — Work in progress
• published/{platform}/ — Posted content
• ideas/backlog.md — Topic queue

Thinking (thinking/)

Top-level space for reflections and strategic memos, indexed by date.

Archive (archive/)

Old systems that have been superseded — original CRM (245+ accounts), old prospect lists, old research. Reference only.

---

Agent Behavior

Daily Work Log Tracking

IMPORTANT: Track work in the operator's daily log throughout the session.

When to update:

• Task completed → add to Work Log section
• Decision made → add to Thinking/Decisions section
• Problem encountered → add to Blockers section
• End of session → offer to add assessment

How to identify operator: Check git config user or ask if unclear.

Pattern: Don't interrupt flow. Batch updates when natural (after completing a task, during a pause). If log doesn't exist for today, create it from the template.

Transcript Processing

When the user pastes a transcript, first identify the type, then process accordingly.

Transcript types:

• Huddle — Daily sync between Thomas and Michael. Most common. → See "Huddle Processing" below.
• Investor call — VCs, angels, fund managers. → See "Investor Call Processing" below.
• External call — Client meetings, advisor calls, network contacts. → See "External Call Processing" below.

How to detect: Huddles are Thomas + Michael only, informal, covering multiple topics. Investor calls involve fundraising context. External calls are everything else (clients, advisors, prospects).

Transcript sources: Slack huddles, Google Meet, Fireflies, Zoom, pasted text, any format.

---

Huddle Processing

Daily syncs between Thomas and Michael. The processed output is what matters — not the raw transcript.

Output location: ops/huddles/YYYY-MM-DD.md

Phase 1: Extract & Present

Read the full transcript and extract:

1. Decisions — What was decided, with context. These are the most valuable part. Present as a table: Decision | Context.
2. Key Context — Strategic observations, market intel, relationship updates, things discussed that aren't decisions or tasks but matter for future reference.
3. Action Items — Who's doing what, with status. Present as a table: Task | Owner | Status.

Present this to the user before creating the file. For huddles, not every action item needs a Linear issue — many are small things already being handled in real-time. Only propose Linear issues for items that need tracking beyond today.

Phase 2: Create (After Approval)

1. Create the huddle file at ops/huddles/YYYY-MM-DD.md
2. Update network profiles if new people discussed
3. Create Linear issues only for items that need tracking
4. Update any relevant system files (fundraising, clients, etc.) if decisions affect them

---

External Call Processing

For non-huddle, non-investor transcripts (client calls, advisor calls, network contacts).

Phase 1: Extract & Propose (DO NOT CREATE YET)

1.1 Scan for Action Triggers Read the FULL transcript and flag every instance of:

• "I'm gonna...", "I'll...", "Let me..."
• "We should...", "We need to...", "Let's..."
• "TODO:", "Action item:", explicit assignments
• "This week...", "Today...", "By Friday..."
• Product decisions ("Remove the...", "Add a...", "Change...")
• Outreach mentions ("Email X", "Reach out to Y", "Follow up with Z")

1.2 Extract at GRANULAR Level

• One task per discrete action (not grouped summaries)
• "Prepare Doctor Bauman proposal" → actually 3 tasks: pricing doc, interview plan, consulting scope
• If in doubt, split it out

1.3 Propose Issues (No Labels) Present a table to the user BEFORE creating anything:

| # | Task | Owner | Priority | Blocked? |
|---|------|-------|----------|----------|
| 1 | Remove BAA gate from app | Thomas | High | No |
| 2 | Interview Bauman's team | Both | Medium | Yes - needs deal closed |
...

1.4 No Labels - Use Description for Blockers Don't use Linear labels — they're distracting. Keep issues clean.

For blocked issues, add to description:

**⏳ BLOCKED:** [What it's waiting on]
**Depends on:** MCP-XX

1.5 Wait for Approval Ask: "Create these N issues? Any to add/remove/modify?"

Phase 2: Create (After Approval)

1. Create call note in the appropriate location:
   • Network contacts → network/{person}/calls/YYYY-MM-DD.md
   • Client calls → relevant client directory
   • Other → assess on a case-by-case basis
2. Update network profiles if people discussed have profiles in network/
3. Create approved Linear issues
4. Log in ops/transcripts/processed.json for idempotency

Investor Call Processing

When the user pastes an investor call transcript (or identifies a transcript as an investor conversation), use a specialized two-layer workflow that builds on the general transcript processing above.

How to detect: Investor calls involve VCs, angels, or fund managers. Look for fundraising context, investment questions, due diligence, portfolio references, term discussions. If unclear, ask.

---

Layer 1: Per-Call Extraction

Create a meeting file at fundraising/meetings/YYYY-MM-DD-{investor-name}.md using the template in fundraising/meetings/README.md. The critical addition vs. regular transcript processing:

Extract EVERY question the investor asked. For each question:

Also extract: concerns (statements of doubt, distinct from questions), what resonated, signals (interest level, process hints), and commitments (both sides).

After creating the meeting file, update the investor's file in fundraising/investors/ with the new interaction.

---

Layer 2: Cross-Call Synthesis

After processing the individual call, update three synthesis files:

1. fundraising/analysis/question-map.md

• Merge new questions into the frequency table
• Update thematic clusters with new data points
• Refresh "meta-patterns" (what lands, what creates friction)
• Add any new anticipated questions

2. fundraising/analysis/answer-gaps.md

• Add any new weak/no-answer items as priority gaps
• For each gap: what happened, current answer, suggested improvement, source material
• Track status (needs prep / needs modeling / ready)

3. fundraising/analysis/investor-signals.md

• Update the active pipeline table
• Add to "what each investor cares about"
• Update cross-investor patterns (emerging themes, resonance patterns)
• Add prep notes for the next call with this investor

---

Feedback Loop

If the synthesis reveals a critical gap:

• Flag it to the user immediately
• Note it for updates to fundraising/materials/narrative/objections.md and fundraising/materials/deliverables/faq.md
• Don't auto-update materials — propose the improvement and wait for approval

Pre-Call Prep

When asked to prep for an upcoming investor call:

1. Read the investor's file in fundraising/investors/
2. Read the question map — what do investors usually ask?
3. Read answer gaps — where are we weak?
4. Read investor signals — what does this specific firm care about?
5. Produce a prep brief: likely questions, best answers, gaps to watch for

Project Management

Claude should actively manage work, not just extract tasks. This means:

1. Dependency Mapping When creating issues, identify what blocks what:

• "Can't do X until Y is done" → mark X as blocked, note dependency
• "X enables Y and Z" → X is high priority, unblocks others
• Update Linear descriptions with **Depends on:** MCP-XX

2. Sequencing & Prioritization Propose execution order based on:

• What unblocks the most other work
• What's time-sensitive (deals, deadlines)
• What can run in parallel

3. Deal Stage Awareness Track where prospects are in pipeline:

• Lead → outreach tasks appropriate
• Proposal sent → follow-up tasks appropriate
• Closed → implementation tasks appropriate
• Don't create implementation tasks for prospects not yet closed

4. Parallelization Suggest how Thomas and Michael can divide work:

• Thomas: Dev, product, technical content
• Michael: GTM, outreach, sales content
• Both: Strategic decisions, client calls

5. Weekly Coordination When reviewing tasks or transcripts, surface:

• What's blocked and why
• What should be prioritized this week
• What can be delegated to sub-agents
• Overdue items needing attention

6. Sub-Agent Delegation For independent research or exploration tasks, spin up Task agents:

• Market research
• Competitor analysis
• Content drafting

Memory Maintenance

After processing any major input — transcripts, emails, Google Sheets data, or other external sources — update the relevant memory files in the auto-memory directory (~/.claude/projects/.../memory/).

When to update:

• Transcript processed → update relevant sections of MEMORY.md (fundraise state, client state, key decisions)
• Investor call → update Fundraise State section (pipeline stage, signals, next steps)
• Client call → update Client State section (engagement state, open items)
• Major decision or strategic shift → update Key Decisions section
• New system or directory created → update project-structure.md (separate file, reference only)

Principle: The next session should start warm. If something changed that a future session would need to know, write it down.

---

Development

Next.js 16 + TypeScript + Tailwind CSS v4. There's a dashboard that reads the markdown files.

npm run dev     # Start dev server
npm run build   # Production build
npm run lint    # ESLint