Building an HTML Report Generation System from Scratch

The Problem

Every session with an AI coding assistant starts from scratch. The assistant has no memory of what happened in the previous session — no context, no progress, no decisions. This leads to repetitive explanations, lost work, and frustration.

The root cause? No startup protocol. The AGENTS.md file (the configuration file read at every session start) lacked any hook to load previous context, create today's session file, or handle crash recovery.

The Approach

I designed a three-layer solution:

1. Session Startup Protocol

A mandatory startup sequence embedded in AGENTS.md that runs at the beginning of every session:

• Reads active-session.json to identify the current day and previous session
• Loads the previous day's session file into context
• Creates today's session file with the correct structure (Morning/Afternoon time slots)
• Verifies all components are loaded before proceeding

2. Incremental Persistence

To prevent loss from crashes, connection drops, or model switches:

• Every 3 actions (or ~15 minutes), the session state is persisted to disk
• The source of truth is always on disk, never in memory
• Crash recovery reads the last persisted state and resumes from there

3. Consistent Report Templates

Session reports were being generated with inconsistent styling. I created a proper template system:

• A single HTML template with well-defined CSS variables (--primary, --secondary, --accent, etc.)
• A structured grid layout for consistent section placement
• Standardised sections: Executive Summary, Timeline, Key Decisions, Files Modified

Technical Details

The template uses CSS custom properties for theming:

:root {
    --primary: #1E2761;
    --secondary: #CADCFC;
    --accent: #F96167;
    --text: #333333;
    --muted: #666666;
    --success: #4CAF50;
    --warning: #FF9800;
}

And a 12-column CSS grid for layout flexibility:

.container {
    display: grid;
    grid-template-columns: repeat(12, 1fr);
    gap: 8px;
    max-width: 1200px;
    margin: 0 auto;
}

Key Results

• Zero context loss going forward — every session picks up exactly where the last one left off
• Consistent report styling across all generated reports
• Crash recovery that limits data loss to at most 3 actions
• Model-switch proof — switching between AI models doesn't break the session
• 5 role-specific workflows designed: Educator, Student, Content Creator, Data Analyst, PDC Coach

Files Created

File	Purpose
AGENTS.md (updated)	Session startup protocol at the top
2026-05-07.md	Today's session file with proper structure
Templates/Session Report Template.html	Consistent report template
0mem/SKILLS/session-report-workflow.md	Standard Operating Procedure

Key Takeaways

1. Design for failure. The system assumes crashes will happen and handles them gracefully.
2. Disk is truth. Memory is ephemeral; the filesystem persists. Always write state to disk.
3. Consistency reduces cognitive load. A template removes design decisions so you can focus on content.
4. Startup protocols prevent amnesia. A few seconds of setup saves hours of re-explanation.