mehdiforoozandeh

Pick the model tier for every subagent before spawning it, so cheap work runs on cheap models: Opus for anything that writes, runs, implements, designs, or carries consequence; Sonnet for read-only work like research, reading, and summarizing; a Haiku swarm…

Rapid-fire conversation mode: reply in 2–3 sentences max — usually one, sometimes two words — the way a colleague answers across a desk. A hard ceiling, not a target: the reply is sized to the question, and when the whole answer doesn't fit, the part that…

Test whether the user's mental model still matches work that already exists — code the agent wrote, a PRD, a plan, a document tree — usually because they delegated it, skimmed it, or were away while it was built. Derives a tree from the artifact, then asks…

Interactive multiple-choice teaching drill on a subject the user doesn't know yet. Requires a grounding source (the user names one, or it asks, or it searches out a credible one), shows the syllabus for approval, then climbs a general → specific knowledge…

Resolve a stuck trade-off — "we can't improve A without hurting B" — by finding a structural fix that dissolves the conflict instead of trading along it. Names the single element that must be two things at once, then looks for the escape across four…

Re-situate something you (the agent) just said when it drifted into jargon or internal framing the user never grounded — NOT by dumbing it down (ELI5), but by filling in the missing context: a brief backdrop, how it fits together, and where it lands in *this*…

Brainstorm a consequential, no-ground-truth question with a panel of diverse minds, then resolve it dialectically. A casting step derives orthogonal thinking-axes for the question (each a quality to maximize) and casts one persona per axis; the personas…

Set up, compile, preview, and troubleshoot LaTeX in VS Code with the LaTeX Workshop extension — an Overleaf-equivalent local/remote workflow (build-on-save, PDF preview, SyncTeX, latexmk, bibtex citations, LTeX grammar, Live Share). Use whenever working with…

An agentic research companion — a scientific-method lab notebook for a research program. Organize work as a tree of Questions (what we don't know) and falsifiable Hypotheses (testable leaves), each with pre-registered verifiables and findings; a deterministic…

Evolve crux itself — add a capability or fix a recurring flaw in the crux tool, end to end: ideate → build → validate → ship. Turns a feature idea or a "crux keeps doing X" annoyance into a signed-off PRD, a tests-first implementation, a hard validation gate…

Write the Requirements Document for a crux node — the design detail that does not fit, and should not fit, inside a 400-word question or hypothesis. An RD is one immutable design document per node, living in `rd/` beside the tree, linked from the node,…

Turn one crux anchor node and everything under it into a minimal, professional, self-contained HTML slide deck — the "here's what we found" update a research group gives — where every number on a slide is an address into the vault, not a photocopy of one, and…

Launch and manage the crux cockpit — the read-only browser GUI over a crux research vault (`crux serve`). This skill runs it beginning to finish: locate the vault, start the server fresh on localhost, verify it actually serves, and hand the user one clickable…

A literature wiki for a crux research vault — Andrej Karpathy's LLM-wiki pattern applied to a crux project. The PI curates immutable sources under `raw/`; you (the agent) compile them into a persistent, interlinked `wiki/` of background, prior methods, SOTA,…

Design and run Karpathy-style autoresearch harnesses in CANDI/EpiDenoise. Use when planning FAFO loops, writing prepare.py/train.py/program.md, scope-fenced agent experiment search, or sandbox/autoresearch/* tags. Not for full sandbox production runs — those…

Analyze CANDI JEPA runs. Triggers on: lejepa/* metrics, encoder_eff_rank, cos_sim_ctx_tgt, sigreg_loss, adaLN_gamma_norm, meta_sens_runtype, SIGReg lambda, pred_hidden bottleneck, UMAP/PCA encoder geometry, encoder collapse, trivial solution, metadata…

Analyze sandbox.train run logs deterministically and act as autoresearch co-researcher on EpiDenoise sandbox runs. Use when the user asks to inspect, compare, rank, or diagnose sandbox runs, mentions metrics.jsonl / resolved_config.yaml / SLURM job IDs / W&B…

Maintain the EpiDenoise/CANDI sandbox experiment idea hub. Use when planning, staging, running, documenting, or interpreting sandbox experiments, hypotheses, B/E/I experiment IDs, idea_*.md files, EXPERIMENTS.md, META.md, or sandbox run findings.

Produce cross-run synthesis docs for sandbox experiments. Use whenever rolling up findings across multiple sandbox runs (sweeps, ablations, head-isolation comparisons, longitudinal F* findings), creating or editing files matching…