Shackleton Framework: Pivot Failing AI Projects Fast
Detect sinking AI plans with 3 traps and a 2-minute diagnostic prompt. Use 4-phase framework—acknowledge ice, inventory survivors, excavate real mission, rebuild from wreckage—with 5 copy-paste prompts to turn dead projects like GREENHOUSE v1-2 into v4 in one evening.
Spot Sunk AI Projects via 3 Psychological Traps and Binary Diagnostic
AI projects fail when builders ignore divergence between plan and reality, mistaking activity for progress—like Shackleton's crew pumping water from the crushing Endurance for 10 months, dooming nothing while all 27 survived after letting it sink. Three traps lock you in:
- Plan attachment: Persist with outdated specs despite changes (Kahneman & Tversky's commitment bias; evidence: continuing to original whiteboard plan after 400 lines of vibe code).
- Sunk cost paralysis: Weigh past time/money/emotion retrospectively instead of future outcomes prospectively (behavioral economics distinction; sunk costs are gone—calculate only forward value).
- Vehicle confusion: Obsess over the tool (e.g., agent architecture) vs. problem (ready-to-hand vs. present-at-hand; symptom: pitching system before outcome).
Run this diagnostic prompt on your project for direct diagnosis (pick 1-3 or combo, with evidence):
I'm going to describe an AI project... [paste traps] Here's what's going on: [DESCRIBE YOUR SITUATION HONESTLY]
Core pivot question (binary, no hedging): If rebuilt from scratch now, would you build the same thing? Yes = fix 2-3 breaks. No = plan dead, proceed to triage. Author's GREENHOUSE agent (idea-tendering AI) hit this: v1-2's two commands (/plant, /signal) forced user sorting, adding patches worsened UX; one-entry rebuild in one evening succeeded.
Excavate Survivors and Real Mission to Rebuild Leaner
Post-diagnostic, inventory via two columns (prompt pushes forgotten items):
- SURVIVED: Problem understanding, audience needs, research/context files, taste criteria, failure lessons (these transfer across pivots).
- SANK: Architecture, file structure, tools, specific implementation.
GREENHOUSE survived: idea-tending core. Sank: rigid commands.
Then excavate real mission by iterating past surface answers (prompt chains 'why does that matter?' to irreducible goal):
- Wrong: 'Content sorting system.'
- Better: 'Stop losing ideas.'
- True: 'Create conditions for ideas to grow autonomously vs. filing cabinet death.'
This reveals mission obscured by building (Shackleton's: reach pole, not preserve ship).
Generate Rebuild Brief from Wreckage for One-Evening Wins
Final prompt drafts 1-page brief:
- Real mission (1 sentence).
- Carry-forward assets (specific, e.g., context files).
- Leave-behind (with why).
- Simplest v1 plan.
- 1 lesson new plan must respect.
Outcome: Smarter, leaner systems (GREENHOUSE v4 simpler/faster). Foundation (structured .md files, thinking) outlives implementations—build assuming first version sinks. All 5 prompts in RobotsOS; start with Phase 1 tonight on stalled project.