RPI Skills: Research, Plan, Implement
CompletedI had a version of this workflow sitting in a skills folder since May with no frontmatter and none of the machinery that lets it survive across threads. In July I tore it down and rebuilt all three skills plus a README.
The one idea that changed everything. The handoff doc is the real product of each phase. Research does not hand the next phase its 40 files of notes. It hands over roughly 200 lines. Plan reads only that. A fresh thread reading one 200-line doc starts at 10 to 15 percent context usage instead of dragging 60,000 tokens of research chatter forward. That is where the savings live, and the quality goes up too, because the failure mode you are avoiding is context pollution.
Where your attention actually pays. Reviewing 400 lines of specs beats reviewing 2,000 lines of generated output. Research review matters most, plan review second, and you can skim the build reports. Spend the editing minutes at the front.
Every phase needs a "done when" the agent can check itself. Otherwise "looks done" is the only signal, and "looks done" is not a signal. It does not have to be a test suite. "The sheet reconciles to the export" counts. "The doc answers these four questions" counts. Then a fresh subagent reviews it adversarially, because fresh eyes have no attachment to the work.
Subagents do the noisy reading. The main thread never reads 40 files. And a small state file per project makes any thread resumable, which matters more than it sounds when you come back three weeks later.
How I tested it: three live agent runs, including a planted veto buried in a handoff doc and a planted contradiction in a plan. The mechanics caught both. Six ambiguities surfaced and got patched before I packaged it.
Open question I haven't answered: is 200 lines the right handoff cap for bigger projects, or does it need a per-project override? Still unsure.
Adapted from HumanLayer's frequent intentional compaction and Anthropic's Explore, Plan, Code guidance. Shipped as installable skill cards.