Skip to content
mubaroq edited this page Aug 17, 2026 · 2 revisions

FAQ

Baca dalam bahasa Indonesia

Do I have to install all six?

No. Each skill works fully on its own. Install what you need. If a neighbour is present some steps go deeper — but nothing stalls because another skill is absent.

How do I invoke them?

You do not. A skill activates by itself when you mention something relevant — "help me write the introduction", "my manuscript was rejected without review", "I want to do a systematic review".

Do I need Python?

Not for the five text skills. Without Python some dimensions are done manually and the skill is required to say that coverage is reduced — rather than skipping quietly. If you do have it, no pip install is needed: every script in those five is stdlib-only.

visualisasi-data is different, because it actually draws. It needs matplotlib and numpy (pip install matplotlib numpy). Without them the drawing stops and says so — it does not quietly degrade to a worse chart — while the half that chooses the visual form, being prose, keeps working.

Why does visualisasi-data break the stdlib-only rule?

Because the rule was never the point; not surprising you was. A skill that renders figures cannot render them without a rendering library, and pretending otherwise would mean shipping a skill that fails at the moment you need it.

So the dependency is stated up front rather than discovered mid-figure, and it is kept to the smallest set that does the job: matplotlib and numpy, both required; scipy, pillow, pypdf, and mne optional, each deepening one specific check. Notably topomap does not need MNE.

Can it draw a figure for a field that is not in its domain list?

Yes, and that is deliberate. Fourteen domain files are shortcuts, not the limit of coverage. The visual substrate is set by the structure of the data, not by the name of the field, and there is a finite number of structures — a separate reference maps 22 of them to their substrates, with a five-step procedure for working out the convention of an unfamiliar field.

The rule is explicit inside the skill: never fall back to a bar chart just because the field is not listed.

All my references suddenly came back UNVERIFIED. Is my manuscript broken?

Almost certainly not. Suspect the CA certificates first — Python from python.org on macOS does not ship them, and the result is that every lookup fails at once. See Requirements.

A network failure is not a manuscript finding.

My manuscript is in Indonesian. Does that work?

Yes. The conversation follows your language — ask in Indonesian, get Indonesian. The manuscript's language is separate and follows the target journal — usually English, and there is a dedicated check for Indonesian calques that commonly slip through: "It is known that…", "It can be concluded…", 0,05 where 0.05 was meant.

My corpus is national journals. Scopus returns very few hits.

That is a feature of index coverage, not of the field. A re-measurement on 12 August 2026 found indexed Indonesian journals (OJS) deposit abstracts to Crossref on 92–100% of records, while three Elsevier education journals deposited 0%.

slr-cowork includes Garuda, Moraref, SINTA, and PTKIN repository routes, plus a rule requiring you to state coverage limitations in the Limitations section.

I do not speak Indonesian. Can I use these?

Yes, fully. The skills reply in whatever language you write in — ask in English, get English.

Both this documentation and the Indonesian version are maintained: English pages live in docs/, Indonesian in docs/id/.

The one thing still in Indonesian is the content of the skill files themselves. That text is instruction to the model, not to you — Claude reads it and answers in your language. You never need to read it.

May I use these for paid training?

Not without permission. The licence is CC BY-NC 4.0 — non-commercial. For research and teaching (including regular university classes) no permission is needed at all; just credit the source. See Licence.

May I modify and redistribute my own version?

Yes, for non-commercial purposes, with attribution. NOTICE.md must travel with it — there is one at the root and one inside each skill, and their contents differ according to what that skill carries.

Can I use these in ChatGPT / Gemini?

The files are ordinary markdown, so any model can read the content. But the Skill format — automatic loading, progressive disclosure, script execution — is Claude-specific. Elsewhere you would have to paste the content manually.

I found a methodological error. Where do I report it?

Open an issue in the repo. Include the skill, the file, and the source you believe is correct — with a DOI if there is one, it is far faster to verify.


← Back · Installation · Workflow · Requirements · Licence

Clone this wiki locally