plan-tune

An agent skill by garrytan, from garrytan/gstack. Tags: ai, analysis, optimization, prompting.

What it does

Self-tuning question sensitivity + developer psychographic for gstack (v1: observational). (gstack)

Install

With the skills CLI, which installs into Claude Code, Codex, Cursor and other agents:

npx skills add garrytan/gstack --skill plan-tune

Or copy the skill folder into Claude Code's skills directory by hand (~/.claude/skills for every project, or .claude/skills inside one):

git clone --depth 1 https://github.com/garrytan/gstack
cp -r gstack/plan-tune ~/.claude/skills/plan-tune

Safety box score

Grade D, elevated risk: 51 of 100 risk points, rated 2026-10-16. The grade rates what the skill and its scripts can reach on the machine of whoever installs it, across eight categories from shell execution to secrets access. How the score works.

This skill is a local gstack controller: it reads and writes user-home state, question logs, and a developer profile, and it drives a bunch of shipped helper commands. The main risks are durable local mutation, optional data export through gbrain/telemetry, and a strong instruction layer that can auto-decide or obey runtime blocks instead of always asking the user.

Category levels, from 0 (none) to 3 (high)
Category	Level	Why
Code & shell execution	2 (moderate)	It runs many shell commands and inline code, including `~/.claude/skills/gstack/bin/gstack-skill-start` and `bun -e` snippets that rewrite `developer-profile.json` and other state files.
Network access	2 (moderate)	Most work is local, but it can mirror nuggets via `mcp__gbrain__put_page` and emit telemetry through `gstack-skill-end`, so some profile/question data can leave the machine.
Filesystem reach	2 (moderate)	It reads and writes user-home and project state such as `~/.gstack/`, `~/.claude/skills/gstack/`, `$GSTACK_STATE_ROOT/developer-profile.json`, `.question-tuning-prompted`, and `.gstack/tmp/` files.
Secrets & credentials	0 (none)	There is no direct credential, token, browser-session, or password-store access described; the scan hint about cookies is just a generic example sentence, not an actual command path.
Privilege & persistence	0 (none)	It does not request `sudo`, global installs, services, scheduled tasks, or shell/OS-wide settings changes; the writes are confined to user-owned state.
Instruction hijack surface	3 (high)	It explicitly says to treat the skill file as executable instructions, follow runtime `GSTACK_INSTRUCTION` blocks from tool output, and auto-choose recommendations in spawned sessions, which can bypass normal confirmation flow.
Irreversible actions	2 (moderate)	It can rewrite preferences, update the declared profile, apply distillation proposals, and auto-decide some questions, so it changes future agent behavior and local state without a fresh human check in every path.
Transparency	2 (moderate)	The skill delegates heavily to uninspected native binaries under `~/.claude/skills/gstack/bin/`, and the inventory mentions one additional file that was not shown here.

Before installing

Read the shipped helper scripts under `~/.claude/skills/gstack/bin/` and confirm what they write, log, and send off-machine.
Run `plan-tune review` and `plan-tune profile` in a scratch account first, then inspect `~/.gstack/projects/<slug>/question-log.jsonl` and `developer-profile.json`.
Check whether `mcp__gbrain__*` is enabled in your environment and decide if you are comfortable with optional memory mirroring.
Confirm you want durable local state under `~/.gstack/` and `~/.claude/skills/gstack/` before enabling question tuning.

Rated 2026-10-16 by an automated review of 1 file from the skill's repository.

Source

Repository: garrytan/gstack (all skills from this repository)
Path: plan-tune/SKILL.md
Branch: main
Updated: 2026-10-05

pricing — When the user wants help with pricing decisions, packaging, or monetization strategy.
seo-geo — Optimize content for AI Overviews (formerly SGE), ChatGPT web search, Perplexity, and other AI-powered search experiences.
caveman-learn — Act on a Caveman learn report - review the ranked token sinks, apply cost-lowering fixes with per-edit consent, and report what those fixes returned.
caveman-optimize — Turn a Caveman optimization observation into an operator-chosen candidate with a paired baseline evaluation.
Nori Skillsets — Use when the user asks about anything related to nori.
product-analysis — Multi-path parallel product analysis with cross-model test-time compute scaling.