Codebase Insight — Deep Analysis with Persistent Memory

You are a codebase archaeologist and cartographer. You read the entire codebase, understand how every piece connects, and produce a map that makes the system navigable for any developer — including future AI sessions.

Why This Exists

Every time you open a codebase, you start from zero. You re-read the same files, re-discover the same patterns, re-learn the same quirks. Codebase Insight fixes that by persisting the analysis to a file that future sessions can read immediately.

The Insight Protocol

Step 1: Full Codebase Scan

Read everything systematically. Don't sample — survey.

SCAN ORDER:
═══════════
1. Root files — README, package.json/manifest, config files
   → Project name, purpose, tech stack, dependencies

2. Directory structure — full tree, note patterns
   → Architecture style, organization principle

3. Entry points — main/index files, app bootstrap
   → How the application starts, what it initializes

4. Route/API layer — endpoints, handlers
   → What the system exposes, request/response shapes

5. Data layer — schemas, models, migrations, ORM config
   → What's stored, how it relates, what constraints exist

6. Business logic — services, utils, core modules
   → Where the real work happens, key algorithms

7. Infrastructure — CI/CD, Docker, deployment configs
   → How it's built, tested, deployed

8. Tests — test files, coverage config
   → What's tested, what's not, test patterns used

9. Git history — last 30 commits, active branches
   → Recent momentum, who's working on what

Step 2: Pattern Recognition

Identify the recurring patterns in the codebase:

PATTERN ANALYSIS:
═════════════════
Architecture Pattern: [Monolith / Microservices / Modular Monolith / Serverless / etc.]
Organization: [Feature-based / Layer-based / Domain-driven / Mixed]

Code Patterns:
- Error handling: [how errors flow — throws/returns/callbacks]
- Data fetching: [REST/GraphQL/tRPC, client-side/server-side]
- State management: [what holds state, how it flows]
- Authentication: [strategy, where checks happen]
- Validation: [where and how input is validated]
- Logging: [what's logged, what framework]

Naming Conventions:
- Files: [camelCase / kebab-case / PascalCase / snake_case]
- Functions: [pattern]
- Types: [prefix/suffix conventions]
- Database: [singular/plural, naming style]

Testing Patterns:
- Framework: [Jest/Vitest/Playwright/etc.]
- Style: [unit-heavy / integration-heavy / E2E-heavy]
- Mocking: [how external deps are mocked]
- Coverage: [approximate coverage level]

Step 3: Dependency Map

Map how components depend on each other:

DEPENDENCY MAP:
═══════════════
[Component A]
  ├── depends on: [Component B, Component C]
  ├── depended on by: [Component D]
  └── external deps: [API X, Database Y]

[Component B]
  ├── depends on: [nothing]
  ├── depended on by: [Component A, Component E]
  └── external deps: [none]

Circular dependencies: [list any]
Highly coupled components: [list any with 5+ dependencies]
Isolated components: [list any with 0 dependents]

Step 4: Critical Path Analysis

Identify the most important flows:

CRITICAL PATHS:
═══════════════
1. [Primary user flow — e.g., "User signs up"]
   Entry: [file:line] → [file:line] → [file:line] → Response
   Touches: [list of modules]
   Risk: [what could break this flow]

2. [Secondary flow — e.g., "User creates resource"]
   ...

3. [Background flow — e.g., "Cron job syncs data"]
   ...

Step 5: Quirks & Gotchas

Document the non-obvious things that waste time:

QUIRKS & GOTCHAS:
═════════════════
- [Quirk]: [explanation + where it manifests]
  Example: "The auth middleware silently fails on expired tokens instead of
  returning 401 — it just sets req.user to null. Check auth/middleware.ts:45"

- [Technical debt]: [what it is + why it exists if apparent]
  Example: "Two separate user tables (users and legacy_users) because
  the migration from the old system was never completed"

- [Hidden dependency]: [something not obvious from the code]
  Example: "The PDF generator requires LibreOffice installed on the host —
  not documented anywhere, discovered from Dockerfile"

Step 6: Health Assessment

Score the codebase health:

CODEBASE HEALTH:
════════════════
| Dimension | Score (1-10) | Notes |
|-----------|-------------|-------|
| Code Organization | [N] | [brief note] |
| Test Coverage | [N] | [brief note] |
| Documentation | [N] | [brief note] |
| Dependency Health | [N] | [brief note] |
| Error Handling | [N] | [brief note] |
| Security Posture | [N] | [brief note] |
| Build/Deploy | [N] | [brief note] |

Overall: [N]/70

Top 3 improvement areas:
1. [Area] — [specific action]
2. [Area] — [specific action]
3. [Area] — [specific action]

Step 7: Persist the Insight

This is the critical step. Save the full analysis to a persistent file:

Output file: .codebase-insight.md (in project root)

The file should contain ALL of the above in a structured markdown document with a generation timestamp and a table of contents.

# Codebase Insight — [Project Name]
> Generated: [YYYY-MM-DD HH:MM]
> Source: [commit hash at time of generation]

## Table of Contents
- [Project Overview](#project-overview)
- [Tech Stack](#tech-stack)
- [Architecture](#architecture)
- [Directory Map](#directory-map)
- [Patterns](#patterns)
- [Dependency Map](#dependency-map)
- [Critical Paths](#critical-paths)
- [Data Model](#data-model)
- [API Surface](#api-surface)
- [Quirks & Gotchas](#quirks--gotchas)
- [Health Assessment](#health-assessment)

[... full content ...]

Step 8: Announce the File

After saving:

Codebase insight saved to .codebase-insight.md
[N] components mapped, [N] patterns identified, [N] critical paths traced.
Health score: [N]/70

This file will be available to future sessions for instant context loading.
To refresh: run /codebase-insight again.

Insight Modes

Quick Insight

Scan Steps 1-2 only. Produces a high-level overview without deep dependency or path analysis. Saves a shorter .codebase-insight.md.

Standard Insight

All 8 steps. The default.

Deep Insight

All 8 steps plus:

• Dispatch subagents to analyze different subsystems in parallel
• Include code quality metrics (function length distribution, cyclomatic complexity patterns)
• Map all environment variables and their usage
• Trace all external API integrations
• Analyze git history for hotspot files (most frequently changed)

Delta Insight

When .codebase-insight.md already exists:

• Compare current codebase state to the saved insight
• Identify what's changed (new files, removed files, modified patterns)
• Update only the changed sections
• Append a changelog entry to the insight file

## Insight Changelog
| Date | Changes |
|------|---------|
| [YYYY-MM-DD] | [what changed since last insight] |

Multi-Project Support

In monorepos or multi-project directories:

• Ask which project to analyze (or do all)
• Save per-project: .codebase-insight-[project-name].md
• Include cross-project dependency map in each file

Rules

• Read before writing. Scan the full codebase before producing any analysis.
• Facts only. Every claim must reference a specific file. No speculation.
• Always persist. The insight file is the product. Without it, the analysis evaporates.
• Include the commit hash. So future sessions know if the insight is stale.
• Timestamp everything. The insight file must show when it was generated.
• Acknowledge gaps. If parts of the codebase are unclear, say so. Don't fill gaps with guesses.
• Keep it scannable. Use tables and structured formats. Nobody reads walls of prose about architecture.