Reset between skills

Run each skill in a fresh context. Why the artifact-on-disk handoff exists, and how to implement multi-phase plans.

The pipeline’s design rests on one operational rule: every skill boots with a clean context window. Between any two skill invocations, run /new (or whatever your harness calls “clear context”) before kicking off the next one.

The artifact-on-disk handoff exists precisely for this reason. Each skill writes its output to .rpiv/artifacts/<stage>/, and the next skill reads it back. The chain is durable, not stateful: your next session can pick up a half-finished run without losing context, because state lives in markdown files, not the chat window.

Skipping the reset is the single most common cause of muddled output.

The reminder pattern

Every skill that hands off to a next step emits a reminder when it finishes:

🆕 Tip: start a fresh session with `/new` first — chained skills work best with a clean context window.

It applies to every transition in the chain, not just the one you just saw. (The lone exception is commit, the chain terminus: it has no next step to remind you about.) The tip is a polite nag; the rule is hard.

When NOT to reset

One case: within a single skill invocation. Several skills run multi-turn work inside one call. discover runs a one-question-at-a-time interview, code-review dispatches parallel specialist agents, and blueprint, research, and explore each run their own internal checkpoint loops. The reset rule applies between skill invocations, never within them.

Every other transition resets. That includes phase-by-phase /skill:implement (see below).

Implementing a multi-phase plan

If your plan has a single phase or is small enough to land in one shot, run:

/skill:implement .rpiv/artifacts/plans/<plan>.md

This executes every phase in a single invocation. One session, one pass, done.

For larger plans with multiple phases, run one phase per session with a reset between every phase:

/skill:implement .rpiv/artifacts/plans/<plan>.md Phase 1

Then /new to reset the context, review the diff, and run the next phase in a fresh session:

/skill:implement .rpiv/artifacts/plans/<plan>.md Phase 2

The plan artifact carries state between phases. It tracks - [x] completion markers as each success criterion lands, so a fresh context per phase keeps the agent focused on the slice in front of it instead of dragging accumulated chat noise from earlier phases.

Next steps

• Pick your path: which chain to run for what you’re shipping
• Handoffs: pause the pipeline and resume it later from a fresh session
• Walk the chain: the mid-size path demonstrated end-to-end