Every working session has a shift change. With a human team, someone explains what shipped, what broke, which decision looked odd but was intentional, and where the next person should start. An AI-assisted development session needs the same thing. It just has a more literal clock.

The context available to an assistant is finite. Reading the repository, following a long plan, inspecting a large file, and doing the actual work all spend from the same budget. When that budget runs out, the next session does not inherit understanding by osmosis. Without a durable handoff, it begins with an archaeological dig.

DaftKit, the portable-skills companion to DaftPlate, treats that limit as an engineering constraint. Its normal loop is /orient, work, then /handoff. Plans are divided into phases that can be completed independently inside a fresh session, and every phase declares the files the next session actually needs to read. Context is budgeted before implementation rather than discovered to be exhausted halfway through it.

/orient creates a compact working brief. It reads the repository instructions, the scaffold provenance, the headers of the code map, the latest changelog entry, current Git state, a small set of open issues, and the most recent handoff. It is looking for shape and recent movement, not trying to memorize the entire codebase. The target is a useful brief under 15,000 tokens, followed by a stop. Orientation is preparation, not permission to wander into the task.

It also knows when not to run. A concrete, self-contained request may cost less to complete than a full orientation pass. Context management is not a ritual; it is deciding which information is worth loading for the work in front of you.

/code-map handles the opposite problem: the file that is too large to read casually. It creates a greppable index of symbols and useful ranges so a future session can locate the relevant function, class, or region without pouring the whole file into context. The map is not a second copy of the code. It is a set of road signs, which is both cheaper and less likely to become another source of truth.

At the other end of the shift, /handoff appends a dated entry to the active plan. It records what shipped, the commits involved, discoveries, open work, and the next phase’s read manifest. It also calls out things the plan got wrong and marks intentional decisions that should not be reverted by a well-meaning future session. Then the note is committed. An uncommitted handoff is just a thought with a short life expectancy.

The plans themselves use the same discipline. A phase estimates the size of its required reading and leaves room for the churn of implementation. If it will not fit comfortably inside roughly 150,000 tokens, the work splits at a real boundary such as a toolchain, repository, or dependency seam. The answer is not to pretend the ceiling is taller.

This improves more than AI output. A new developer benefits from the same code map. A reviewer benefits from the written decision trail. A project manager benefits from phases with explicit entry and exit conditions. The assistant’s constraint simply makes a familiar delivery problem impossible to ignore: continuity has to be designed.

There is overhead. Someone has to maintain the code map, write the handoff, and keep a phase manifest honest. On a tiny change, that ceremony would cost more than it saves, which is why orientation can step aside for self-contained work. On a project that crosses sessions, the alternative is paying for rediscovery again and again, with interest charged in inconsistent decisions.

AI can write a remarkable amount of code during one shift. The measure of the system is whether the next shift can understand why it exists, verify what happened, and continue without starting over.