Requirements

Requirements & environment

Baca dalam bahasa Indonesia

For the five text skills, nothing is mandatory. They work without any of the tools below — what shrinks is speed and traceability, not the obligations. The skill is required to say what it skipped rather than skipping it quietly.

visualisasi-data is the exception, and it is stated plainly in its own row: a skill that draws cannot draw without a drawing library.

Tool	For	Without it
Python 3	mechanical sweeps, fidelity gate, PRISMA audits	done manually, reduced coverage
pandoc	reading .docx manuscripts	export the manuscript to .md or .tex
MCP scholar / zotero	citation verification + retraction detection	falls back to WebSearch/WebFetch, then to manual flagging
R + the robvis package	risk-of-bias traffic-light figures	the robvis web app (no R needed), or a study × domain table
matplotlib + numpy	drawing figures in visualisasi-data	that skill cannot render — its design guidance still applies in full

Python

Every script in the five text skills is stdlib-only — no pip install, no virtualenv. Tested on both the macOS system Python 3.9.6 and 3.12.

python3 -V     # check

Not installed? macOS: xcode-select --install. Windows: winget install Python.Python.3.12.

⚠ CA certificates on macOS

Python from python.org does not ship CA certificates. The consequence is that verify_refs.py reports every reference as UNVERIFIED — including the ones that genuinely exist.

That is a network failure, not a manuscript finding. If all references turn UNVERIFIED at once, suspect the certificates first:

python3 -c "import ssl,os; p=ssl.get_default_verify_paths().openssl_cafile; print(p, os.path.exists(p))"
"/Applications/Python 3.12/Install Certificates.command"   # adjust the version number

Each Python version has its own certificate directory — installing a new version reintroduces the problem.

pandoc

Only affects .docx input. For .tex and .md, every script runs fully.

brew install pandoc          # macOS
winget install JohnMacFarlane.Pandoc

MCP for citation verification

Without MCP, verification still runs through WebSearch/WebFetch — resolve the DOI at doi.org, match title, first author, year, and journal name. What cannot be done without MCP: automatic retraction detection.

The rule that holds at all three tiers: a citation is never accepted because it "looks plausible". A plausible-looking author–year–journal combination is the signature pattern of a fabricated reference.

matplotlib + numpy (visualisasi-data only)

The one place in this repository where pip install is genuinely required:

pip install matplotlib numpy

Everything the skill draws is plain matplotlib — no seaborn, no plotly, no R. In particular topomap does not need MNE: the 10-20 electrode coordinates are baked into the module. Set prefer_mne=True only if you want them read from MNE directly.

Four packages are optional and each deepens exactly one thing:

Optional	Deepens	Without it
scipy	topomap interpolation, ridgeline KDE	coarser interpolation, histogram instead of KDE
pillow	raster checks in figcheck.py	PNG dimension and dpi checks are skipped
pypdf	PDF font-embedding check in figcheck.py	embedded-font check is skipped
mne	reading montages straight from MNE	the built-in 10-20 coordinates are used

Without matplotlib the skill does not fall back to a worse chart — it stops rendering and says so. The half of the skill that chooses the visual form is prose and keeps working: you still get the referent test, the domain route, the figure rules, and a caption. You just draw it yourself.

R + robvis (slr-cowork only)

Only for risk-of-bias figures. There is a web app, so R is not required at all.

install.packages("robvis")

---

← Back · Installation · Workflow · FAQ · Licence