How working with Jig looks
You do not need to learn commands or skill names. You talk to your agent:
You: Add CSV export to the reports page.
Agent: This touches the report pipeline and the download endpoint, so I classify it as
structural (T2): analyze, plan, implement, review, verify. I have created the task
csv-export on its own branch. Here is what I found in the code and the plan …
Behind that answer, the agent picked the jig-task skill, looked at the risk of the change,
chose a route, created a task workspace and loaded only the project knowledge relevant to
reports. The agent usually picks the right skill from what you say; when it does not, name it —
/jig-task in Claude Code, $jig-task in Codex. A one-line typo fix gets a three-step route; a change to authentication gets design, your
approval and an independent review.
What you get
Work sized by risk
Five classes, from a typo to a change that touches security or money. The route grows with
the risk, never with habit.
Knowledge that stays true
Terms, architecture, rules and decisions live in
.ai/knowledge/ next to the code. Every task
ends by deciding what it taught the project.You decide where it matters
At a human gate the agent shows you the design itself — not a summary — and waits.
Your existing docs, adopted
Rules you already keep in
docs/, your own ADRs or another tool’s instruction files are linked
in place, not copied.Where to start
New project
Start from an idea: specification, architecture, rules, then the first tasks.
Existing project
Bring Jig into a codebase you already have, and adopt the documentation it keeps.