AI Compass / Knowledge cards

Project map instead of rereading

Claude Code with a code map

Checked 17 October 2026 Michael Engel Costs Prompting

A knowledge graph of the code is a map: it shows which parts exist and how they connect. Tools such as Graphify or Trail-code (formerly Graft) generate it on your computer from the source code. Claude Code then reads the map instead of opening file after file in every session.

What it is for

In large projects, the agent finds the right place faster and uses less context. As a side effect, you see for yourself how your project is built.

Steps

First clarify whether it is worth it. According to Jack Roberts, only from about 500 files. For a small project, Claude Code’s normal search is enough.
Install Graphify, according to the README with uv tool install graphifyy, then graphify install. The package name has a double “y”. According to the README, other packages with a similar name are not part of the project. graphify install detects the assistant and sets up the skill.
In Claude Code, call /graphify .. According to the README, the code is parsed locally, without a language model. It only calls a model for documents, PDFs and images. This creates the folder graphify-out/ with the overview GRAPH_REPORT.md.
Ask questions and name the Graphify skill explicitly. Then ask whether the map was really used.
The alternative with hooks is Trail-code (MIT license). According to the vendor’s site, you install it with npm install -g @trailhq/trail-code and start trail init in the project. In Claude Code, hooks then attach matching locations to every prompt. In the video the commands are still called graft init and graft build; according to the README they still work.
Run a comparison: the same task with and without the map, noting time and context use. Neither video shows a measurement method.
After major refactoring, check whether the map is still accurate and regenerate it if necessary. An outdated map leads the agent to the wrong place (my own assessment).

Limits

The savings figures come from the vendors: Jack Roberts speaks of “70x”, Trail-code claims 60% less time and 32% lower cost over 162 runs, and Graphify advertises its cloud offering with 80% fewer tokens. AI LABS tested Graft itself and measured a small gain: 39 instead of 47 minutes, 31 instead of 35% context.
According to the README, Trail-code writes its hooks into ~/.claude/settings.json, so they apply to all projects. On trail init it reads your previous Claude Code sessions in this repo once. Anonymous usage statistics are switched on and can be disabled with trail telemetry disable or DO_NOT_TRACK=1. According to the README, trail init --dry-run shows beforehand what would be written, and --no-global leaves out the global entries. Read what the hooks do first.
Graphify has moved since the video. The repo is now at Graphify-Labs/graphify, and the old address redirects. The local part is free according to the README (Apache 2.0); there is also a paid cloud offering.
With third-party packages, double-check the name before installing. Fake packages with similar names are a known attack route.

Sources

Jack Roberts: Claude Code + Graphify = Insane Agentic OS, 8 June 2026, 3:13–5:23. Self-promotion for course and community
AI LABS: Github Top Trending Tool Just Fixed The AI Agent’s Biggest Problem, 8 Sept 2026, 3:34–4:21, 8:07–11:53. Sponsored by Hedra
Jack Roberts: Top 5 Claude Code Skills… 100,000+ github stars, 25 Apr 2026, 3:14–3:46. Short links with suspected affiliate use (Firecrawl, Zapier)
Graphify Labs: graphify, README and PyPI package graphifyy, retrieved 17 Oct 2026
Trail: Trail-code and README, retrieved 17 Oct 2026

Status

Checked 17 Oct 2026 · Maturity new (tools available, not tested by me; savings from vendors and one creator test only)

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab