agentleFS
Sign inStart free
LuisUrrutiaLuisUrrutia/skills/humanize/SKILL.md

humanize

Skill2 starsChanged 15 days agoMITView on GitHub

Use only when the user explicitly asks to humanize text or to run it through humanizer ("humaniza esto", "humanize this", "pásalo por humanizer"). Not for ordinary rewriting or editing requests.

  • Reads credentials

We split this skill into the rules it gives an agent

1 rule. 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 63 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Before handing the rewrite back, compare it with the original yourself: names, dates, quotations and claims must survive, and nothing may …

    line 52

“Common” is measured: kept by at least 6 in 10 of the 63 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Humanize

Rewrite the text with scripts/humanize from this skill's directory (also on PATH as humanize). It runs Luis's humanizer model, a fine-tuned Gemma 4 12B, with a local llama-server when the machine has the memory for it: 32 GiB of unified memory on Apple silicon or 16 GiB of CUDA VRAM for Q8_0, or 16 and 12 GiB for Q6_K. Otherwise it calls the Runpod Serverless endpoint, which needs RUNPOD_API_KEY and HUMANIZER_URL (the endpoint's /completion URL), read from the environment or, when missing there, from fish universal variables. stderr names the backend used. When a variable is missing or the script exits 1, report its message as the blocker; do not rewrite the text yourself instead.

humanize draft.md > draft.humanized.md
printf '%s' "$TEXT" | humanize
  • Languages: the model only handles English and Chinese. For any other language, tell the user and stop; do not translate to work around it.
  • Input: pass only the prose to rewrite. The script keeps fenced code blocks and one-line Markdown headings verbatim and splits long text into pieces of about 3,000 characters at paragraph breaks.
  • Local runs: the first local run downloads the GGUF once (12.7 GB for Q8_0) into $XDG_CACHE_HOME/humanize/, or ~/.cache/humanize/ when that variable is unset, resuming an interrupted download; later runs load it in seconds. When the machine qualifies but llama-server is missing or cannot load the model, the script falls back to Runpod and says why on stderr; tell the user, since Runpod is billed. Run the first local run in the background: the download takes minutes. --backend local or --backend runpod forces one backend.
  • Runpod cold start: the endpoint scales to zero. The first call after idle waits 6 to 40 seconds for a worker, or several minutes when Runpod places it on a host without the cached model; the script retries for up to 7 minutes. Run it in the background when the harness would time out sooner.
  • Cost: local runs are free. On Runpod, generation dominates: about USD 0.02 for 2,000 words, plus under 0.01 per cold start, billed to Luis's Runpod balance. Send each text once; do not loop for a better sample unless the user asks.

Check the result

Exit code 2 means a piece still lost or gained a number or URL, or was cut off at the output limit, after one automatic retry; stderr names it. Exit code 0 checks only numbers, URLs and truncation.

Before handing the rewrite back, compare it with the original yourself: names, dates, quotations and claims must survive, and nothing may be added. Sampling runs at temperature 1.0, so tone can drift (slang, emoticons, a changed register). Fix a factual slip by restoring the original wording in that sentence; report tone problems instead of rewriting the text yourself. Give the user the rewrite and a list of every fix or warning. Never send or publish the rewrite on their behalf.

Endpoint

The endpoint is a Runpod Load Balancing endpoint running llama.cpp's llama-server with humanizer-12b-Q8_0.gguf from jialinyyzz/humanizer-GGUF on a 24 GB GPU, scaling from 0 to 1 workers, with the repository as its cached model. If calls keep timing out, check it with the runpod:endpoint-ops skill: Runpod sets max workers to 0 after 7 days without requests.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free