A reusable work-item harness for projects that build fast and cut hard: write the change down before you build it, prune the living set aggressively, and keep the whole record append-only. The name is the crux — pruning here means retire without erasing.
For running real changes through a project — especially when a lot of the work is done with AI agents — without losing the thread of what is open, what shipped, and why. Code or not: it is disciplined Markdown plus a method.
This is the harness extracted from two live projects (a persistent-AI substrate and the city it forked from), shipped here empty by design: the method and the templates, with the worked examples left in the repos that actually used it.
Write the change down before you build it. Review it against a rubric frozen ahead of the work. When
it ships or dies, move it to history/ with a one-line evidence note — append-only, nothing deleted.
So the folder is always an honest snapshot of what is actually open, and the archive is a real record
rather than a graveyard of half-truths.
majors/— large arcs (multi-file, new subsystems, behavior changes). Shape inmajors/MAJOR_SCHEMA.md.minors/— bounded changes. Shape inminors/MINOR_SCHEMA.md.history/— the append-only archive: shipped and retired items, each with its evidence. It is gitignored and mirrors the structure one level down (history/majors/,history/minors/). The full ledger is kept locally; only live work is published. (One convention across every project using this kit.)harness/— the method: operating model, work-item lifecycle, agent execution protocol, quality gates, git/release policy, a pruning playbook, and copy-pastetemplates/. Start atharness/00-ADOPTION_GUIDE.md.
- Copy this tree (or just
harness/+ the two schemas) into your repo, e.g. underimprovements/. - Read
harness/00-ADOPTION_GUIDE.mdandharness/01-OPERATING_MODEL.md. - Draft your first item from
harness/templates/(MAJOR_ITEM_TEMPLATE.md/MINOR_ITEM_TEMPLATE.md). - Keep
majors/andminors/to live work; archive tohistory/as you close things.
It works for non-code projects too — the unit is a decision you can review and a result you can check, not a diff.
This scaffold ships empty on purpose. The same harness, fully populated with real worked examples
(triaged, scrubbed, kept append-only), runs the public improvements/ ledgers of the projects it came
from — look there to see what a year of actual items looks like.
Pairs naturally with a memory discipline (durable artifacts kept apart from prunable interpretations, behind a single auto-loading standing brief). Two halves of the same honesty: this keeps your intentions reviewable; a memory workspace keeps your understanding reviewable.
MIT. Built by Levi Banks with AI collaborators across two live projects; the method is the contribution. Take it and run.