Prototyper Docs GitHub ↗

Context System

The 95-file / 624KB AI context library at .opencode/context/ is the project’s shared working memory. It follows three rules:

  1. MVI — Minimum Viable Information. Core concept + 3–5 bullets + minimal example + reference.
  2. Function-based folders — concepts/, examples/, guides/, lookup/, errors/.
  3. Navigation at every level — navigation.md per category, with quick routes and loading strategy.

This page is the short version. The full source is in Context → Core → Context System → Overview.

Core principles

1. Minimal Viable Information (MVI)

Extract the minimum an AI agent needs to use a concept:

  • Core concept (1–3 sentences)
  • Key points (3–5 bullets)
  • Minimal example (<10 lines)
  • Reference link to full docs

Goal: scannable in <30 seconds. Reference full docs, don’t duplicate them.

2. Concern-based structure

Organize by what you’re doing (concern), then by how (approach / tech). Two patterns:

Pattern A — Function-based (for repository-specific context):

category/
├── navigation.md
├── concepts/        # What it is
├── examples/        # Working code
├── guides/          # How to do it
├── lookup/          # Quick reference
└── errors/          # Common issues

Pattern B — Concern-based (for development context, spans multiple techs):

category/
├── navigation.md
├── {concern}/       # Organize by what you're doing
│   ├── navigation.md
│   └── {approach}/  # Then by approach/tech

3. Token-efficient navigation

Every category and subcategory has a navigation.md with:

  • ASCII tree for quick structure scan (~50 tokens)
  • Quick routes table for common tasks (~100 tokens)
  • “By concern / type” sections (~50 tokens)

Target: 200–300 tokens per navigation file. Faster loading, less cost, quicker AI decisions.

4. Self-describing filenames

Filenames tell you what’s inside:

  • code.md → code-quality.md
  • tests.md → test-coverage.md
  • review.md → code-review.md

No need to open a file to understand what’s in it.

5. Knowledge harvesting

Extract valuable context from AI summaries / overviews, then delete them. The workspace stays clean, the knowledge persists.

The current top-level layout

.opencode/context/
├── navigation.md                  # Top-level entry point
├── core/                          # Universal standards & workflows
├── prototyper/                    # Prototyper project context (Tauri v2 + React 19)
├── project-intelligence/          # Domain-specific project intelligence
├── development/                   # Generic development references
└── ui/                            # UI patterns (Tailwind v4, shadcn, React, animations)

The Context section in the sidebar is auto-generated from this directory on every build.sh run — see Context → Overview for the live tree.

File size limits

File type Max lines
Concept files 100
Example files 80
Guide files 150
Lookup files 100
Error files 150
README files 100

Why: forces brevity. If you need more, split into multiple files or reference external docs.

Validation checklist

Before creating or updating a context file:

  • Core concept is 1–3 sentences?
  • Key points are 3–5 bullets?
  • Example is <10 lines of code?
  • Reference link is included?
  • File is <200 lines total?
  • Can be scanned in <30 seconds?

If any answer is “no”, apply more compression.

What next