"The butterfly counts not months but moments, and has time enough." — Rabindranath Tagore

A skill for Claude Code, OpenCode, GitHub Copilot CLI, OpenAI Codex CLI, Gemini CLI, and Goose. Run /tagore on the AI-drafted prose you just generated; it rewrites the text to sound human and scores how close it got. Named in homage to Rabindranath Tagore, whose prose carried what frontier models reach for and miss: a point of view, specificity over abstraction, and restraint over puffery.

Tagore attacks slop in two directions at once:

• Removes the tells. 29 cataloged AI patterns get scrubbed: significance inflation, em dash overuse, vague attributions, AI vocabulary, rule of three, copula avoidance, and more.
• Adds what's missing. Point of view, stakes, specificity, restraint, varied rhythm, and trust in the reader.

The skill runs every rewrite through a six-stage pipeline ending in an 8-dimension scoring rubric (5 mechanics + 3 substance) with a 56/80 pass threshold and subtotal floors so a piece can't pass on mechanics alone.

Tagore is a synthesis of two complementary skills:

humanizer is a deep catalog with a pipeline. stop-slop is a tight rulebook with a scoring gate. Tagore fuses them into a single workflow and adds three substance dimensions (Specificity, Restraint, Voice) that neither skill scored on its own.

curl -fsSL https://raw.githubusercontent.com/apurvrdx1/tagore/main/install.sh | bash

The installer detects every supported harness on your system and installs Tagore into each. It also drops a universal symlink at ~/.agents/skills/tagore/ so any agent honoring the cross-tool convention can find it.

curl -fsSL https://raw.githubusercontent.com/apurvrdx1/tagore/main/install.sh | bash -s -- --platform claude

Supported --platform values: claude, opencode, copilot, codex, gemini, goose, agents (universal).

git clone https://github.com/apurvrdx1/tagore.git && cd tagore && ./install.sh --all

Pick your harness and copy the tagore/ directory (with SKILL.md inside) into the matching location:

The universal ~/.agents/skills/ path is recognized by OpenCode, Copilot CLI, and Codex, so installing there once covers three harnesses.

/tagore Humanize this draft: [paste text]

Or invoke the skill explicitly via the Skill tool. Provide a writing sample inline for voice calibration:

/tagore Humanize this. Use my style from this sample: [sample]. Now rewrite: [draft]

@tagore Humanize this draft: [paste text]

Voice calibration works the same way: paste a sample with your prose alongside the draft.

/tagore [paste text]

After copying the skill into the directory, run /skills reload (or restart the CLI) so Copilot picks it up. Verify with /skills info tagore.

codex
> /tagore Humanize this draft: [paste text]

Codex auto-loads everything under ~/.codex/skills/; no reload step required.

Invoke by name and paste the text. Same six-stage pipeline runs regardless of the harness.

0. (Optional) Voice calibration from sample
1. Draft rewrite — apply the 8 core principles, scrub the 29 patterns
2. Pre-delivery checklist — 12 mechanical yes/no checks
3. Score 1–10 on eight dimensions (5 mechanics + 3 substance, revise if < 56/80)
4. Self-audit — "What makes this still obviously AI generated?"
5. Final rewrite incorporating the audit
6. (Optional) Brief change summary

Pass threshold: 56/80 (70%), with no subtotal failing on its own (Mechanics ≥ 35/50, Substance ≥ 21/30).

A piece can pass mechanics and fail substance. That's the "clean but soulless" failure mode the substance dimensions specifically catch.

Plenty of "anti-slop" prompts exist. Tagore's distinguishing features:

1. Quantitative gate. 8 dimensions, 80 points, 56 to pass. Forces a numerical decision instead of vibes.
2. Two failure modes named explicitly. Inflated slop (puffery) vs. flattened slop (passive narrator-from-a-distance). The rubric tests both.
3. Self-audit loop built into the process. "What makes this still obviously AI generated?" → revise. Catches what the catalog scrub missed.
4. Voice calibration from a sample. Match the user's actual writing instead of producing generic "human" output.
5. Adds, not just removes. The Personality and Soul section prescribes what to put in (opinions, stakes, mixed feelings), not just what to take out.

See examples/essays-with-scores.md for full before/after rewrites with scoring breakdowns.

MIT. See LICENSE.

Built from two excellent prior works:

• humanizer by blader, based on Wikipedia:Signs of AI writing (WikiProject AI Cleanup).
• stop-slop by Hardik Pandya.

Both are MIT-licensed; this merge is too.

Issues and PRs welcome. The most valuable contributions:

• New AI tells the catalog misses (with before/after examples)
• Edge cases where the rubric scores wrong
• Voice calibration improvements for non-English text
• Translations of the skill into other languages

If tagore saved you from sounding like a chatbot in a doc that mattered, drop a coffee. The catalog grows from real submissions, and your support buys time to read them, score them, and ship them in the next release.

"LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases." — WikiProject AI Cleanup

Good writing is rarely the most statistically likely thing. That gap is what Tagore tries to close.