gstack: 98% Less Context Per Prompt Than The /investigate Skill.
Injected size
TOKENS CARRIED PER PROMPT2% of the context the skill costs to carry.
Turns per task
GBRAIN, CLAUDE CODE454 runs, every one passed.
gstack is Garry Tan's skill suite for coding agents. It runs on Claude Code with gbrain as its memory. A skill is a general manual loaded whole into context on every prompt. /investigate is 15,593 tokens. Memorable learned the same job from one session and handed back a 293-token procedure.
A 15,593-Token Manual, Loaded Every Time.
On the same three-bug fixture the skill added 55% input tokens and changed turns 0%. A general manual applies to novel work. On a task that recurs it costs context and changes nothing.
One Procedure, 293 Tokens.
One claude -p session solved the fixture cold. gbrain's session-end hook captured the trace. The extraction API turned it into a 293-token procedure: files touched, the command that proved the fix, exit codes, in order.
gstack skill
GENERAL MANUAL15,593 tokens in context on every prompt, needed or not.
Written for any task, so it cannot say where this one lives.
0% change on the fixture. Input tokens up 55%.
Learned procedure
PROCEDURAL MEMORY2% of the skill's context. Same model, same fixture.
The files touched last time and the command that proved the fix, with its exit code.
Held on an independent replication.
Session-End Hook, Extraction, Recall.
Session-end hook
When a Claude Code session finishes, gbrain's SessionEnd hook hands the session to Memorable. The transcript never leaves; only the prompt and allow-listed tool arguments do.
293 Tokens. Turns Down 19%.
| Metric | Skill | Learned procedure | How it was measured |
|---|---|---|---|
| Tokens carried per prompt | 15,593 | 293 | Same fixture, same model |
| Turns per task | 0% change | Down 19% | The skill added 55% input tokens for no change in turns |
| Turns, receipted replication | 16 | 13 | Two independent runs, every run passed |