x-psychology-transformer-gnn-workflow

X Psychology Transformer + GNN Workflow

Use this skill when the user wants to turn recent X posts into a small deep-learning analysis that treats feature scores as psychological signals and then uses a Transformer + GNN to cluster or interpret the posting style.

Purpose

• model the user's posting style as latent psychological axes
• keep the analysis lightweight enough to run on a consumer GPU such as RTX 3060
• ground the analysis in real public post artifacts rather than invented examples
• produce a reproducible report and a file-backed result

Inputs

• recent public X post history, usually from ~/.hermes/lm-twitterer/activity.jsonl
• a compact feature scoring scheme such as:
  • utility
  • proof
  • contrast
  • self-relevance
  • fork-differentiation
  • execution-first

Recommended interpretation axes

Map the above into a small psychological space such as:

• agency: willingness to act, ship, and execute
• epistemic grounding: reliance on evidence, repo status, results, or proofs
• opposition: contrast, tension, hype-vs-reality framing
• self_reference: how much the post connects to the reader or the speaker's self-positioning

Workflow

1) Collect data

Load the latest public posts from the activity log. Keep only real public posts; avoid drafts, secrets, or private artifacts.

2) Convert posts to feature vectors

Turn each post into a fixed-length vector of feature scores. Use a simple heuristic scoring layer first; the workflow is meant to be lightweight and reproducible.

3) Derive psychological targets

Create small latent targets from the features. Keep the mapping transparent and explainable.

4) Train a Transformer

Use a Transformer encoder over the feature tokens for per-post representation learning.

5) Train a GNN

Build a similarity graph over posts and run a tiny GCN/GCN-like model over the learned embeddings.

6) Cluster and interpret

Cluster the learned embeddings and summarize each cluster with:

• mean feature profile
• mean psychological profile
• representative post
• a plain-language interpretation

7) Save a report

Write a JSON report and a Markdown implementation log in the repo _docs folder.

Operational notes

• Prefer CPU only for tiny smoke tests; use CUDA when available for the real run.
• For Windows mother-ship PCs, check nvidia-smi and then verify torch.cuda.is_available() inside the project venv.
• If the project venv lacks pip, bootstrap it with python -m ensurepip --upgrade before installing wheels.
• If scikit-learn or other analysis deps are missing in the venv, install them explicitly rather than assuming the base environment has them.
• Keep the dataset small enough to run quickly on a consumer GPU.

Output format

Return:

• device / CUDA status
• number of posts used
• transformer loss
• GNN loss
• cluster summaries
• one short psychological interpretation

Pitfalls

• Do not claim clinical or diagnostic meaning from the axes.
• Do not use private or secret data as training input.
• Do not overfit the interpretation to a tiny sample; label it as exploratory.
• If the GPU is visible to the OS but not to PyTorch, install a CUDA-enabled PyTorch wheel in the active venv and rerun.

Relationship to other skills

Pairs well with:

• x-post-feature-scoring-workflow
• hermes-x-posting-workflows
• ebbinghaus-memory

Use this skill as the analysis layer when the user wants to go beyond scorecards into a small learned model.