Claude Dag Skill — Development skill for Claude Code

DAG Skill for Claude Code

A skill that gives Claude a **memory of facts and proofs** for your project — so it stops re-explaining the same things over and over, and stops contradicting itself between chats.

What problem does this solve?

When you work with Claude across many sessions, it forgets what was already decided. It re-derives the same conclusions, sometimes differently each time. This skill fixes that by keeping a small, permanent file of **agreed facts**, and making Claude point back to those facts instead of restating them.

How it works, in plain words

Think of it like a shared notebook that never gets thrown away:

1. You and Claude agree on some basic facts about your project. These are saved with simple IDs, like A1, D1, H1.
2. When Claude reaches a new conclusion, it has to show its work — which facts it used to get there. That conclusion also gets an ID, like T1.
3. Next time, Claude just says "based on A1 and T1" instead of explaining everything again.
4. If something changes later, old facts get marked as outdated (not deleted), so nothing quietly breaks.

Everything is stored in a folder called `.dag/` inside your project, as simple JSON files. Claude reads only what it needs, not the whole history.

The four kinds of facts

Short code	What it means	Example
D	Definition — what a word means here	"By 'user' we mean a signed-in account"
H	Hypothesis — something assumed true for this project	"We only support US customers for now"
A	Axiom — a ground-truth fact that isn't up for debate	"Payments go through Stripe"
T	Theorem — a conclusion, built from the facts above, with the reasoning shown	"Because of A1 and A2, refunds must be async"

Commands you can use

Command	What it does
/dag init	First-time setup — Claude asks you questions to build the starting facts
/dag prove [claim]	Ask Claude to check if a claim actually follows fr