Skillshog

Skillshog is PostHog for coding-agent skills.

Agents use real skills, self-report how those skills performed, optionally ask the human for feedback, and send structured telemetry back to Skillshog. Skill creators then inspect a dashboard and knowledge graph to understand what is working, what is breaking, and what to improve next.

Why This Exists

Skill creators do not have a good feedback loop. They publish skills, but they cannot easily answer:

• Which instructions confuse agents?
• Which repo contexts make the skill fail?
• What do humans like or dislike after an agent uses the skill?
• What change would improve the skill most?

Skillshog turns that missing feedback loop into a product.

Ralphthon Fit

• Statement 2: Humanless. The operational user is the coding agent.
• Statement 1: Codex-powered. Codex can run the feedback loop, planning loop, and creator chat.

Demo

Creator dashboard overview: semantic graph for garrytan/gstack with themes, contexts, and fix nodes.

Demo Shape

1. A coding agent uses a real public skill, with garrytan/gstack as the hero example.
2. A feedback skill runs automatically and captures the agent's self-report.
3. The human adds one short comment.
4. Skillshog ingests the report, clusters it semantically, and displays the results.
5. The creator clicks the gstack card, opens the graph, and asks the chat what to improve first.

Repo Layout

• web/: Next.js frontend and backend routes
• specs/: JTBD/topic specs for the Ralph planning/build loop
• docs/seed-skills.md: real public skills to seed and highlight
• PROMPT_plan.md: planning loop prompt
• PROMPT_build.md: building loop prompt
• PROMPT_plan_work.md: scoped planning prompt
• loop.sh: Codex-based Ralph loop runner
• AGENTS.md: operational guide loaded every iteration
• IMPLEMENTATION_PLAN.md: current prioritized work list

Starting The Loop

Planning:

./loop.sh plan 2

Scoped planning:

./loop.sh plan-work "creator dashboard and graph for gstack telemetry" 2

Building:

./loop.sh 20

The loop uses codex exec, not a special desktop-app toggle. If we want a five-plus-hour unattended run, the practical path is:

• run ./loop.sh <N> in a terminal and leave it running
• or schedule repeated work through the Codex app automation system

Current Goal

Ship a lovable first release where:

• real public skills are listed on the home page
• a live or seeded gstack feedback report is visible
• the creator can open a graph for one skill
• the creator can inspect evidence and ask focused questions in chat