Galivelop-workflow
If it isn’t pushed, it doesn’t exist: the way we work with AI, packaged for any project.
A versioned workflow for coding with AI agents: handoffs between sessions without losing anything, picking the right skill and model before starting, mandatory review before publishing, and fewer tokens without losing quality.
Working with an AI agent for weeks has a hidden problem: any session can end at any moment (out of tokens, container recycled, another account taking over) and whatever wasn’t written down is lost. On top of that, every project ends up inventing its own rules and copying them by hand, and the copies drift apart.
Galivelop-workflow extracts that method from Tarkor, this portal and other workshop projects, where it already works, and turns it into a versioned repository: rules in AGENTS.md (an open standard any agent understands), plus skills, scripts, hooks and templates in one place. An improvement is made once and reaches every project, and each one declares its own details in a small config file.
It comes in two stages. In the first, an AI agent applies the workflow to a project by reading this repository; that’s where it is today, with version 0.3.0, which already takes in what we learned from the pilots on this very portal, on Tarkor and on its installer. In the second, an installer will do it without AI. It will open to the public only once it has worked for weeks on real projects built with different technologies.
What’s ours and what we reuse, kept apart. The workshop built the method (handoffs, the log, three phases before publishing), the nine skills of our own, the hooks and scripts that enforce the rules, the templates and the way it is applied and updated. For thinking before coding (grilling a plan, research, prototyping, diagnosing, designing modules, the retro) it includes twelve skills by Matt Pocock, unchanged and under their licence: they taught us to work this way.
- Handoffs without losses: A log kept up to date in the same commit as the work, and commit + push at every stable point. When a session opens, or after the context is cleared, a hook returns where things stood: the next session carries on without asking.
- Choose before starting: The selector skill decides which skill and model each task calls for, and orchestrate-dealer hands the mechanical work to clean-context subagents on the cheapest model that does it well.
- Nothing unreviewed reaches main: Three phases before publishing and a review mark in the repository. Git hooks won’t push code without it or with failing tests, and they scan for secrets before every commit.
- Fewer tokens, same quality: Clean-context subagents, capped compaction and the opt-tokens skill, which measures where the spend went and proposes cuts that don’t make the result worse.
- For any project: Each project declares its own settings (language, check, protected branch, what counts as content) in a small file; gw-init applies the workflow and gw-doctor says what’s missing, with the command that fixes it.
- Improves with every project: gw-harvest gathers what each project changes or invents; the good parts come back as a new version and gw-update brings it to every project, showing the changes first.
- Galivelop-workflow: If it isn’t pushed, it doesn’t exist: the way we work with AI, packaged for any project.
- toVeriAI: Truth isn’t binary: a score that explains how far you can trust a news story.
- Tarkor: A narrative RPG where the world moves on even when you stop.