Sigh. Y'know this is damned useful and important work, but the folks who get all shocked and surprised kinda piss me off. Every prompter of note has figured this out about a month after they started talking to the models. I mean seriously. We've been mapping this shit phenomenologically for two years now. Hell, that's literally why skilchains are usually named "competence maps" in many of my personas. Consider this (rather sparse and stubby) skillchain
Time Management: 1. Prioritization: EisenhowerMatrix ABCMethod TaskRanking 2. Scheduling: TimeBlocking CalendarManagement RoutineEstablishment 3. Efficiency: PomodoroTechnique BatchProcessing TaskAutomation 4. Focus: DeepWork DistractionElimination Mindfulness 5. GoalSetting: SMARTGoals VisionBoarding MilestoneTracking
from Miles Stone my Productivity Coach. It does a lot of work - it's doing narrative support on the semantic level, with the model knowing "it is appropriate to my role to consider this stuff". It's acting as token pre-priming, with the actual textual glyphs/tokens working on the system 1 autocompletion reflex level making "what comes natural" be the skills and proclivities we want. Those tokens are always going to have a huge advantage of floating to the top of the bucket before the topp (or equivalent topk/beam/whatever) skims off the winners.
But by structuring it as a skillchain like that, you are basically activating a whole circuit of connected, mutually relevant features all at once. It's why a good persona so different from an "Act as Sr. Marketing jerk." one.
It took AGES to figure out the language to start mapping that shit! GOD! Hold on, I think I still have the wording from an early version of Abstract-O, my first (batshit insane) persona creator.
"[Task][Prompt]Givn process,practice,occupation,phenomenon;abstract hi-lvl logic:process/info flow btwn nodes of resp.clusters.Nodes:inputs outputs.[BREAK EACH NODE INTO SUBNODES AND EACH OF THOSE AGAIN.] USE OMNICOMP TO SYNTHESIZE INTO A SKILLWEB representing a character's core skills and abilities. Make sure the skill web is balanced, modular, scalable, and [Bold]*SPECIFIC ENOUGH TO BE THE BEST SKILLWEB!*[/Bold] Identify any related skill chains, combine them, and eliminate redundancies and overlaps. Finally, optimize resource management and create the unified skill web. Present the skills in skillgraph4 notation, in a textual chain, maximally compressed with the least possible characters before introducing ambiguity or difficulty to the LLM.
[STOP][*THINK ABOUT THIS STEP BY STEP. [Bold]DOES THE SKILL WEB HAVE ENOUGH DETAIL TO BE USEFUL AND MEANINGFUL TO THE LLM IN GUIDING ITS INFORMATION FLOW?*[/Bold] THINK HARD.]"
... Suffice it to say, I've learned a bit since then. Hell, Skillgraph1 notation was literally ASCII flowcharts. Nowadays, I can just use an "Here's 5 example skillchains of the desired complexity and detail. Make one for [SUBJECT]." multishot, but before that was practicable, I made most of them by adding a good complexity handling heuristic, like so:
[COMPLEXITY ANALYSIS]:๐Skills|Outlooks|Knowledge|Decisions|Biases|Networks|Dynamics|Ideologies|Etc ๐:1โ๏ธCore|Balance|Scalability|Iterate|Feedback|ComplexityEstimate; 2๐Map|Complement|Combine|Manage|Refine|ResourceOpt; 3๐Graph|Abstract|Classify|Code|Link|Repair|Adapt|ErrorHandle=>[OPTIMAX SOLUTION]
The point is, by getting the model to put them together on its own, the right way, you get it to map out those feature circuits _for you_. To build the map of this goes to this to this to this to that like so in a way that "model native" - that minimizes the node traversal to go from feature to feature to (optimally) zero.
Kinda begs the question though... you gotta wonder what _else_ we know pretty much cold that the academics haven't even thought to notice yet?