llm-tools

name	llm-tools
description	Use when a session is about to bulk-read 3+ files or any single file over ~400 lines, generate predictable boilerplate (tests, fixtures, config scaffolds, doc templates), or compress a long session transcript — and a cheap OpenAI-compatible worker model (Kimi, DeepSeek, OpenRouter, local Ollama) is available via the llm-ask/llm-write/llm-extract CLIs.

llm-tools — cheap-worker delegation

Overview

Premium-model context is the scarcest resource in a session. These three CLIs route bulk I/O and predictable generation to a cheap long-context worker; the premium model consumes the short answer instead of the raw files. Install + provider config: README.md in this directory (~1-minute setup, any /v1/chat/completions provider).

Routing rules (the rules save the tokens; the CLIs just execute)

Situation	Do this
Would read 3+ files, or one file >~400 lines, to answer a question	llm-ask <files...> -q "..." — read the answer, not the files
Need tests / fixtures / config scaffold / doc template from a spec	llm-write -r <ref> -s "<spec>" -o <path> — then review and edit; never trust blind
Long transcript needs to become doc updates	llm-extract <transcript.jsonl> --sections summary,decisions,todos

Never delegate

• Architectural or design decisions
• Debugging — cheap models miss subtle bugs
• Anything touching auth, payments, PII, deletion, or production data
• Final commits and PR descriptions

If a delegated answer looks wrong or thin, re-read the sources yourself rather than trusting the worker. Worker output is untrusted by default: the worker drafts, the premium model reviews and makes the edits.

Cost reference

Asking a cheap worker to summarize a ~12k-token doc runs ~125× cheaper than the same read on a frontier model, at quality that's fine for "summarize / find facts" — which is why the >400-line / 3-file thresholds pay for themselves almost immediately.

Common mistakes

• Delegating a judgment question because it happens to involve big files — split it: llm-ask extracts the facts, the premium model judges.
• Passing files containing real credentials — redact first; workers don't get secrets.
• Treating llm-write output as done — it ships boilerplate, you ship correctness.