humangarden

Hand-tested

We ran these ourselves.

The directory at large is indexed algorithmically — read once, scored on craft and adoption. The 11 skills below are different: we installed them, fed them real inputs, and kept the artifacts. Each verdict comes from sitting through the run.

The protocol behind every entry below is itself a Claude skill: humangarden-review →

  1. 01

    xlsx

    pandas + openpyxl portion runs clean and the color-coding/formula conventions are sound, but the mandatory recalc.py step hard-crashes without LibreOffice and its error scanner false-positives on text commentary

    human review → agent spec ↗

  2. 02

    slack-gif-creator

    A solid PIL-based GIF toolkit — produced a clean 63 KB looping Slack emoji in 1.6s, validator round-trips, but FPS is silently quantized to GIF-spec centiseconds (15 fps → 16.7) and the 480x480 message variant balloons past 1 MB easily.

    human review → agent spec ↗

  3. 03

    skill-creator

    method doc + small Python toolbox — followed end-to-end it produces a valid runnable skill (SKILL.md + script + example) packaged into a real .skill bundle in under a minute; the bundled scripts only work via `python -m scripts.X` from the skill-creator dir

    human review → agent spec ↗

  4. 04

    pptx

    A pptxgenjs cookbook + python OOXML toolbelt — create-from-scratch path generated a valid 6-slide chart deck in 60ms; visual QA loop needs LibreOffice + Poppler installed

    human review → agent spec ↗

  5. 05

    pdf

    pypdf text extraction works clean on born-digital PDFs; pdfplumber tables silently return false positives on academic layouts — do not trust the fallback path

    human review → agent spec ↗

  6. 06

    internal-comms

    four well-crafted style guides wrapped in a 33-line dispatcher — zero install, format-correct on 3P updates, but silently falls through to "general principles" on incident reports and three other types SKILL.md advertises

    human review → agent spec ↗

  7. 07

    docx

    pandoc + docx-js + an opinionated XML round-tripper, glued by an unusually well-written SKILL.md — runs clean end-to-end in ~1 second on a 3-page memo

    human review → agent spec ↗

  8. 08

    doc-coauthoring

    prompt-only workflow (no code, no install) — a tight three-stage protocol that earns its overhead on substantive docs and silently wastes time on thin ones; Reader Testing via subagents is the real differentiator

    human review → agent spec ↗

  9. 09

    claude-api

    documentation-only reference that Claude reads to write Anthropic SDK code; install path (`pip install anthropic`) works clean, but Python users hit a contradiction — SKILL.md says `output_format` is deprecated yet the SDK's `messages.parse()` (also recommended) still requires it

    human review → agent spec ↗

  10. 10

    canvas-design

    A two-step "write a philosophy, then express it" generator with 54 bundled fonts and zero scaffolding — taste-ceiling skill; rendered an A2 archival poster in 60ms once we picked reportlab

    human review → agent spec ↗

  11. 11

    algorithmic-art

    A p5.js generative-art harness with a strong philosophy-first protocol and a real (if opinionated) HTML template — produces gallery-quality flow-field art in ~4s of headless render time, but the craftsmanship is the author's, not the sliders'.

    human review → agent spec ↗