Learning — Capture and Organize Learnings

Capture valuable insights from coding experiences and save them for future reference. Learnings are stored in .agents/learnings/ as Markdown files with YAML frontmatter.

When to Use

• After debugging a tricky issue — capture the root cause and solution
• When discovering a new pattern or best practice
• After solving a problem that might come up again
• When learning how a library or framework works
• After code review feedback that reveals a new insight
• When making architecture decisions with lessons learned

When NOT to Use

• Creating project-level documentation (use docs skill)
• Writing product requirements (use create-prd skill)
• Recording TODO items or task status (use project-files skill)
• The learning is trivial or easily searchable (just use web search)

Learning Format

Each learning is a Markdown file in .agents/learnings/ with this structure:

---
title: Short descriptive title
date: YYYY-MM-DD
tags: [category1, category2]
context: Brief context (optional)
---

# Learning Title

## What I Learned

[Clear, concise statement of the learning]

## Context

[When and why this was discovered — what problem were you solving?]

## Solution / Pattern

[The actual solution, pattern, or insight — make it actionable]

## Application

[How to apply this in the future — include code examples if helpful]

## References

- Related files: `path/to/file.ts`
- Related commits: `abc1234`
- External docs: [link](url)

Frontmatter Fields

Tag Conventions

Use lowercase, hyphenated tags:

• Language/tech: typescript, go, rust, sql
• Framework: react, hono, prisma, cloudflare-workers
• Concept: debugging, performance, security, testing
• Pattern: error-handling, caching, authentication
• Area: api, database, frontend, devops

Workflow

Option A: Use Helper Scripts (Recommended)

# Create a new learning interactively
bun run scripts/learning-create.ts

# Create with title and tags
bun run scripts/learning-create.ts "Prisma connection pooling" --tag prisma,database

# List all learnings
bun run scripts/learning-list.ts

# List most recent 5
bun run scripts/learning-list.ts --recent 5

# Search learnings by tag or keyword
bun run scripts/learning-search.ts "debugging"
bun run scripts/learning-search.ts --tag typescript

# Show a specific learning (supports partial filename match)
bun run scripts/learning-show.ts 2024-01-15-react
bun run scripts/learning-show.ts --latest

Option B: Manual Creation

1. Create a new file in .agents/learnings/:

   # Use a descriptive filename: YYYY-MM-DD-topic-slug.md
   touch .agents/learnings/2024-01-15-react-usecallback-memoization.md

1. Add the frontmatter and content following the format above

1. The filename should be:

• Date-prefixed for chronological sorting
• Lowercase, hyphenated
• Descriptive of the learning topic

Naming Convention

.agents/learnings/YYYY-MM-DD-descriptive-slug.md

Examples:

• 2024-01-15-react-usecallback-memoization.md
• 2024-01-20-cloudflare-workers-cron-triggers.md
• 2024-02-01-go-context-cancellation-patterns.md
• 2024-02-10-prisma-connection-pooling-gotcha.md

Examples

Example 1: Debugging Learning

---
title: Prisma connection pool exhaustion under load
date: 2024-01-15
tags: [prisma, database, debugging, performance]
context: Production API returning 503 errors under high traffic
---

# Prisma Connection Pool Exhaustion Under Load

## What I Learned

Prisma's default connection pool size (10) is too small for high-traffic APIs.
Under load, requests queue waiting for available connections, causing timeouts.

## Context

Production API was returning 503 errors during peak traffic. Logs showed
"Timeout: Pool connection limit reached" errors from Prisma.

## Solution / Pattern

Increase the connection pool size in your Prisma schema:

datasource db { provider = "postgresql" url = env("DATABASE_URL") relationMode = "prisma" }

generator client { provider = "prisma-client-js" }

Then set pool size via environment variable:

DATABASE_URL="postgresql://...?connection_limit=20&pool_timeout=10"

## Application

- Always set `connection_limit` based on expected concurrent requests
- Monitor connection pool metrics in production
- Consider using PgBouncer for very high throughput

## References

- Related files: `prisma/schema.prisma`, `src/db.ts`
- External docs: [Prisma Connection Management](https://www.prisma.io/docs/orm/more/comparisons/prisma-and-typeorm/connection-management)

Example 2: Pattern Learning

---
title: TypeScript template literal types for API routes
date: 2024-02-01
tags: [typescript, api, patterns]
context: Building a type-safe API client
---

# TypeScript Template Literal Types for API Routes

## What I Learned

Template literal types can enforce correct HTTP method + path combinations at
compile time, preventing invalid API calls.

## Context

Building an API client and wanted to ensure callers only use valid endpoint
combinations (e.g., GET /users, POST /users, but not DELETE /users).

## Solution / Pattern

type HttpMethod = 'GET' | 'POST' | 'PUT' | 'DELETE'; type ApiPath = '/users' | '/posts' | '/comments';

type ValidEndpoint = ${HttpMethod} ${ApiPath};

function callApi<T extends ValidEndpoint>(endpoint: T): Promise<Response> { const [method, path] = endpoint.split(' '); return fetch(path, { method }); }

// ✅ Valid callApi('GET /users'); callApi('POST /posts');

// ❌ Compile error callApi('PATCH /users'); // Type error: not a ValidEndpoint

## Application

Use this pattern when building type-safe clients, routers, or any API
abstraction where method + path combinations matter.

## References

- Related files: `src/api/client.ts`
- External docs: [TypeScript Template Literal Types](https://www.typescriptlang.org/docs/handbook/2/template-literal-types.html)

Example 3: Quick Gotcha

---
title: Bun automatically kills long-running scripts
date: 2024-02-10
tags: [bun, devops, gotcha]
context: CI pipeline timing out
---

# Bun Automatically Kills Long-Running Scripts

## What I Learned

Bun has a default 30-second timeout for script execution. Long-running scripts
will be killed without warning.

## Context

CI pipeline was failing intermittently. Investigation revealed Bun was killing
scripts that ran longer than 30 seconds.

## Solution / Pattern

Use `--timeout` flag or `BUN_TIMEOUT` environment variable:

Extend timeout to 5 minutes

bun run script.ts --timeout 300000

Or via environment variable

BUN_TIMEOUT=300000 bun run script.ts

## Application

Always set explicit timeouts for long-running scripts in CI/CD pipelines.

## References

- External docs: [Bun CLI Reference](https://bun.sh/docs/cli/run)

Managing Learnings

Search by Tag

# Find all TypeScript learnings
grep -l "tags:.*typescript" .agents/learnings/*.md

# Or using the search script
./scripts/learning-search.sh --tag typescript

Search by Keyword

# Search learning content
grep -rl "connection pool" .agents/learnings/

# Or using the search script
./scripts/learning-search.sh "connection pool"

List by Date

# Learnings are naturally sorted by date via filename
ls .agents/learnings/

# Show most recent
ls -r .agents/learnings/ | head -10

Archive Old Learnings

# Move learnings older than 1 year to archive
mkdir -p .agents/learnings/archive/2023
mv .agents/learnings/2023-*.md .agents/learnings/archive/2023/

Best Practices

• Be specific — "Prisma connection pool size" is better than "Database tips"
• Include code — Working examples are more useful than descriptions
• Add context — What problem were you solving? What was the error?
• Tag consistently — Use the same tags for related topics
• Keep it actionable — Future you should be able to apply this without re-reading the original problem
• Date everything — Context decays; dates help prioritize newer learnings

Anti-Patterns

• Don't create learnings for things easily found via web search
• Don't duplicate learnings — consolidate related insights into one file
• Don't use vague titles — "TypeScript trick" is useless; be specific
• Don't skip the "Application" section — that's the whole point

Integration with Other Skills

• After `debug` — Capture the root cause and solution as a learning
• After `code-review` — Record patterns or anti-patterns discovered
• With `explore` — Document findings about unfamiliar codebases
• With `mixture-of-experts` — Save expert insights for future reference