Shackleton Framework: Pivot Failing AI Plans in 4 Phases
When AI projects stall, diagnose with one binary question—'Would you rebuild it now?'—then use 4 phases to inventory survivors, uncover the real mission, and rebuild leaner from wreckage, as proven rebuilding GREENHOUSE agent in one evening.
Three Traps Dooming AI Projects to Sunk-Cost Failure
AI builders cling to failing plans through plan attachment (persisting post-change per Kahneman & Tversky's commitment bias, JSTOR 1738360), sunk-cost paralysis (weighing past investments over future value, ignoring that costs are irrecoverable), and means-end confusion (obsessing over shiny agents like architectures before outcomes). Signals: 'One more tweak,' 'I've spent X weeks,' or describing system specs before problems. Shackleton's crew pumped a crushing Endurance hull for 10 months because the tangible ship felt like progress, despite divergence from the mission—mirroring how AI 'productivity' masks drift.
Copy-paste diagnostic prompt identifies your trap: Describe project, AI flags evidence of 1) plan bias, 2) sunk costs, or 3) tool-love vs. problem-solving.
Binary Diagnostic Splits Fix from Pivot
Core question forces clarity: 'If rebuilt from scratch now, would you build the same?' No hedging—yes means triage 2-3 bugs by severity; no means kill plan. Author applied to GREENHOUSE agent (v1 sorting system with /plant & /signal commands): User bore cognitive load of classifying inputs first, patches bloated it worse. Answer 'no' led to one-entry-point v4 rebuild in one evening—simpler, faster, agent-handled sorting.
Prompt: Describe build, get pushed to yes/no, then targeted fixes or wreckage mode. This separates execution tweaks from directional death, preventing weeks of futile iteration.
4-Phase Shackleton Framework Rebuilds from Survivors
Phase 1: Acknowledge ice—Run diagnostic above.
Phase 2: Inventory survivors—Sort into SURVIVED (problem understanding, audience needs, research, frameworks) vs. SANK (tools, files, architectures). Survivors transfer; e.g., GREENHOUSE kept idea-tending insights, ditched commands.
Prompt: List died plan, get exhaustive columns—pushes forgotten gems like context files.
Phase 3: Excavate real mission—Push past 'built X for Y' via 'Why matter?' chains. GREENHOUSE v1: Not sorting, but 'create conditions for ideas to grow autonomously' vs. filing-cabinet death.
Prompt: State project/purpose, iterate to irreducible goal.
Phase 4: Draft rebuild brief—One-pager: Mission sentence, carry-forward assets, abandon list + why, simplest v1, one key lesson. Builds leaner using wreckage (e.g., Shackleton's lifeboats + instruments for 1,800-mile survival).
Prompt structures it. All 5 prompts in RobotsOS; foundation thinking survives pivots, enabling evening rebuilds vs. days.
Outcomes: Plans for failure upfront—structure files/thinking as transferable. Apply Phase 1 tonight to evening project for mode-shift clarity.