References

The Castle

In the 1997 Australian film The Castle, Dennis Denuto is a small-time solicitor representing the Kerrigan family. Their home is being compulsorily acquired to expand the neighbouring airport. Denuto takes the case to court, but he can’t articulate why the acquisition is wrong. His submission to the judge:

He loses. The family then hires Lawrence Hammill QC for the appeal. Hammill makes the same case — that the acquisition is unjust — but with a structured constitutional argument grounded in Section 51(xxxi) of the Australian Constitution. Same cause, different discipline. He wins.

Denuto’s approach is vibe coding — you sense the answer is in there somewhere, but without structure the output doesn’t hold up. Hammill’s approach is plan-driven development — the same tools, the same cause, but with discipline that makes the result stick.

AI intensifies work

An eight-month study at a 200-person tech company, published in Harvard Business Review in February 2026, found that AI consistently increases work intensity through task expansion, blurred boundaries, and increased multitasking. AI doesn’t reduce work — it intensifies it.

plankit is built for independent developers and small teams working with that reality. Plans narrow a specific task to an approved approach before code is written. Discipline is the multiplier.

Claude’s Cycles

Donald Knuth — author of The Art of Computer Programming and inventor of TeX — published Claude’s Cycles in February 2026. The paper narrates Claude Opus 4.6 solving an open Hamiltonian-cycle decomposition problem from Knuth’s ongoing TAOCP drafts. The proof was subsequently formalised in Lean.

The critical detail: the session ran under explicit plan-driven discipline. Filip Stappers instructed Claude to “update this file [plan.md] before doing anything else” after every exploration, with no exceptions. It was not vibe coding. It was structured AI work producing a real, verifiable mathematical result. Knuth wrote: “It seems that I’ll have to revise my opinions about ‘generative AI’ one of these days.”

Completing Claude’s Cycles

Keston Aquino-Michaels’ Completing Claude’s Cycles (March 2026) picks up where Knuth’s paper left off. Two LLM agents — GPT-5.4 and Claude Opus 4.6 — with a third LLM as orchestrator solve the even case under an identical structured prompt called “Residue.” Heavy formalism — strict logging, a strategy register, periodic synthesis, escalation triggers — compresses Knuth’s 31-exploration odd-case path to five.

The paper’s process line — “The prompt structures process, not thought” — is the formal cousin of plankit’s lighter discipline. Both bet on discipline as the multiplier.

plankit takes these ideas seriously. See what it does.