ProgramAsWeights agent skill

An Agent Skill that teaches your AI coding agent to use ProgramAsWeights (PAW) - compile a natural-language spec into a tiny neural function that runs locally - for fuzzy text tasks like classification, extraction, format repair, fuzzy matching, log triage, and routing.

Install

npx skills add programasweights/skills

Works with Claude Code, Cursor, Codex, GitHub Copilot, and 65+ other agents (the skills CLI installs it into each one's skill directory). Add -g to install globally for every project:

npx skills add programasweights/skills -g

Or try it without installing:

npx skills use programasweights/skills@programasweights | claude

What it does

Once installed, the skill activates automatically when you ask your agent to do fuzzy text -> text work that regex can't handle but a full LLM call per item is overkill for:

• Classify / categorize (sentiment, urgency, intent, topic, spam, ALERT vs QUIET)
• Extract fields from messy text (emails, names, dates, IDs)
• Repair / normalize formats (broken JSON, dates)
• Fuzzy / typo-tolerant matching and near-duplicate detection
• Log triage and intent routing

The agent will check the PAW Hub for a ready-made function, or compile one from a spec with examples, iterate it against test cases, and reuse it locally.

What's inside

skills/programasweights/
  SKILL.md          # the skill (trigger + workflow); the eval loop is embedded inline
  LICENSE.txt       # per-skill MIT license
  references/       # api, writing-good-specs, browser-sdk, troubleshooting (loaded on demand)
assets/             # brand icon (paw.svg, icon-512.png)
SECURITY.md         # exactly what the skill does + data flow

The installed skill is script-free - Markdown instructions only, no bundled executable code. (The Codex/Claude plugin + marketplace manifests in this repo are packaging only and are not yet listed in any directory.)

Manual install (without the CLI)

Copy skills/programasweights/ into your agent's skills directory, e.g.:

• Claude Code: ~/.claude/skills/programasweights/
• Codex: ~/.codex/skills/programasweights/
• GitHub Copilot: .github/skills/programasweights/

Data flow and trust

• Compile sends your spec to the hosted PAW API (https://programasweights.com) and returns a program id. Do not put secrets in a spec.
• Inference runs locally through the official programasweights SDK and works offline after the first download. No user data is sent at inference time.
• Auth is optional (anonymous use works). MIT licensed; read every file before installing.
• Full details, including what the skill does not do, are in SECURITY.md. For reproducibility, pin to a release tag or commit and re-review before upgrading.

Links

• Website: https://programasweights.com
• Agent guide (always current): https://programasweights.com/AGENTS.md
• Docs: https://programasweights.readthedocs.io

Citation

If you use this skill in research, cite Program-as-Weights for the underlying programming paradigm. Also cite Compile by Training when using the training compiler.

Program-as-Weights: A Programming Paradigm for Fuzzy Functions

NeurIPS 2026 · Preprint

@inproceedings{zhang2026programasweights,
  title = {{Program-as-Weights: A Programming Paradigm for Fuzzy Functions}},
  author = {Zhang, Wentao and Hotsko, Liliana and Kim, Woojeong and Nie, Pengyu and Shieber, Stuart and Deng, Yuntian},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems},
  year = {2026},
  url = {https://arxiv.org/abs/2607.02512}
}

Compile by Training: Turning Natural-Language Specifications into Local Neural Functions

EMNLP 2026 System Demonstrations · Preprint

@inproceedings{deng-etal-2026-compile,
  title = {{Compile by Training: Turning Natural-Language Specifications into Local Neural Functions}},
  author = {Deng, Yuntian and Nie, Pengyu and Shieber, Stuart},
  editor = {Agrawal, Sweta and Abad, Alberto},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
  month = oct,
  year = {2026},
  address = {Budapest, Hungary},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2026.emnlp-demos.78/},
  doi = {10.18653/2026.emnlp-demos.78},
  pages = {830--839},
  ISBN = {979-8-89176-461-3}
}