The Problem

Developers pairing with AI agents ship working code but lose the reasoning behind key decisions. A week later, a question like "why a service worker over a content script?" gets no answer because the rationale lived in a chat log that no longer exists. Generic tutorials don't help—they don't know what you built, which choices mattered, or what you already understand.

What This Does

learn-from-session is a Claude Code skill that turns your last AI coding session into a credibility-checked reading path. The SKILL.md file defines a six-step loop: extract 3-6 real topics from the conversation, find 2-3 practitioner sources per topic, run a credibility pass across Hacker News/Reddit/Twitter, recap what you built, quiz you conversationally, and order reading by what the quiz revealed.

The output is a structured markdown report (shown in README.md and assets/example-output.png): a 30-second recap, quiz results per topic, an ordered reading path with credibility evidence, and a "skip for now" list.

How It Is Wired

This is a single-file skill, not a codebase. SKILL.md contains the entire logic as instructions with YAML frontmatter, following the Agent Skills spec. The README documents two install paths: npx skills add notacp/learn-from-session via the skills CLI, or a direct curl of SKILL.md into ~/.claude/skills/learn-from-session/.

What each file is responsible for:

  • SKILL.md — the complete skill definition; all behavior lives here
  • README.md — usage documentation, install commands, trigger phrases
  • LICENSE — MIT license
  • assets/example-output.png — sample output for the README
  • assets/social-preview.png — repository social preview image

Entry point and flow: The skill triggers when you say "I want to learn what we just built" at the end of a session. The agent reads SKILL.md, executes the six-step loop against the session transcript, and produces the markdown report. There is no code to execute, no API calls, no database—the entire "runtime" is the agent following the skill's instructions. Network effects occur when the agent searches for sources and checks credibility, but those are agent actions, not code paths.

The wiring beyond the skill file has not been mapped—there is no code to trace. The blast radius is entirely within SKILL.md; changing the prompt logic changes all output behavior.

How To Use It

Setup (from README, verbatim):

npx skills add notacp/learn-from-session

Or, without a CLI:

mkdir -p ~/.claude/skills/learn-from-session
curl -fsSL -o ~/.claude/skills/learn-from-session/SKILL.md \
  https://raw.githubusercontent.com/notacp/learn-from-session/main/SKILL.md

Configuration: none—no environment variables, keys, or config files. The skill reads your session context directly.

Running it: at the end of a working session, say any of: "I want to learn what we just built", "Find blogs about what we did today", "Go through our chat and get me reading material", or "Quiz me on this session".

Real-World Use

You spend an afternoon building a Chrome extension with Claude Code. At the end, you say "I want to learn what we just built." The skill extracts topics like "Chrome MV3 service worker lifecycle", finds practitioner sources, checks their credibility on HN, recaps why you chose a service worker over a content script, quizzes you on the code you wrote, and orders reading so shaky-but-load-bearing topics come first. Total first pass: under 2 hours.

Code Health & Issues

Static analysis findings (from the pipeline, not opinion):

  • Med/SDLC - No test files detected — untested code paths, repository-wide. Expected for a skill definition, but there's no automated way to verify the skill behaves as documented.
  • Med/SDLC - No CI/CD pipeline detected — no .github/ or CI config. No automated gate for changes to SKILL.md.

The analysis also reports: no lockfile, no committed secrets, and an MIT license present.

The Bottom Line

A pragmatic, well-documented skill that solves a real problem—retaining the reasoning behind AI-assisted decisions. The single-file design is simple to audit and modify, and the install paths are clean. The lack of tests and CI is acceptable for a prompt-based skill but means changes are verified only by manual use. Worth adopting if you regularly pair with Claude Code and want to convert sessions into durable understanding.