camera-lens-travel-eval

STORM-based research skills (storm-research one-shot, costorm-session interactive) plus an example eval skill, for Claude Code / Cowork.

Interactive, human-in-the-loop Co-STORM research: a steerable, turn-based discourse among LLM experts and a moderator over web sources, growing a shared mind map, then synthesizing a cited report. Resumable.

Insert /breakpoint context-checkpoint markers that divide a session into logical blocks so /neuralize can later evict irrelevant context. Slash-invoked only (model auto-invocation disabled).

Perspective-driven, outline-first long-form research (Stanford STORM): multi-perspective research conversations, an evidence-grounded outline, a sectioned cited article, and a deterministic citation audit.

Explain unfamiliar code, systems, or domains by orienting in plain English, scoping blast radius, and walking concrete data flows before diving into implementation detail. Repo-agnostic.

Decompose a design doc into a parallelizable DAG of tasks and run them across git worktrees with an interactive visual workflow. Repo-agnostic.

Test-driven development with a red-green-refactor loop, plus references on deep modules, interface design, mocking, and refactoring.

Report token usage, wall-clock time, and per-model cost for Claude Code sessions from the on-disk ~/.claude/projects JSONL transcripts — grouped by session or git worktree, with a compare mode. Dedups streamed/re-serialized usage and prices per model so totals reconcile with ccusage. Read-only.

Scaffold a repo-coupled multi-agent team (lead / architect / backend / qa / reviewer / devops) tuned to a target repo's conventions — emits project-local agents, slash commands, and a coordination protocol, then adversarially verifies the generated claims.

Selectively evict irrelevant context blocks and reload only what matters (/neuralize) — the surgical alternative to clearing a whole session. Pairs with breakpoint. Slash-invoked only.