dsh-research

From nothing to a talk

Five steps, about twenty minutes, fifteen of which are waiting for dependencies. You end with a deck you can present and a routine you can repeat.

  1. Install DeepSeek Harness

    Needs Node 22.19 or newer. Install it globally rather than running it through npx: a stable version matters more than always having the newest one.

    npm install -g @deepseek-ai/dsh

  2. Install the research plugin market

    Install this one first. It adds a Research plugins page to Settings, and everything else installs from there without going back to a terminal.

    dsh plugin --profile web add dsh-research

    If you already know what you want, take the whole workbench in one paste instead: dsh plugin --profile web add dsh-research dsh-ai4scholar dsh-slides dsh-paper-slides dsh-academy

  3. Add the literature tools and a key

    Run dsh web, open 127.0.0.1:3080, go to Settings → Research plugins, install dsh-ai4scholar, then press Restart dsh in the panel.

    When it comes back, go to Settings → Plugins → AI4Scholar and paste a key from ai4scholar.net.

    The key goes into the dsh credential store — never into the conversation, never to the model. arXiv, bioRxiv and DOI full text need no key, so you can skip this and come back.

  4. Your first search

    Ask in plain language. The guidance forbids answering from memory, so the agent searches before it asserts.

    “Find the strongest evidence from the last three years on your topic, sorted by citations, and tell me what each one concludes.”

    The answer ends with the credits this turn spent and what is left. The free sources — arXiv, bioRxiv, DOI full text — are not billed.

  5. Read it, then present it

    Pick one and have it read the full text — the PDF body, not the abstract. Then ask for the talk.

    “Turn 10.1038/s41586-021-03819-2 into a 12-minute conference talk.”

    It budgets the slot for that venue, builds the deck around the figures, writes the talking into speaker notes, and produces a self-contained HTML deck — presents offline, S reveals the notes, Ctrl+P prints to PDF. Defenses and conference talks also get an editable .pptx.

What else it does

Four research skills come with the pack and load themselves when the task calls for them.

🔍

Pre-submission review

Five reviewer personas audit the manuscript on their own axis and end in a prioritized revision list.

✍️

Introductions

Background, the turn toward the gap, the contribution, signposting, then a polish pass.

📐

Formatting

Reference styles, three-line tables, abstract structure, and reformatting for a different journal.

🔎

Reference audit

Cross-citation matching, DOI checks, journal names, duplicates and self-citation rate.

Common questions

The install printed peer dependency warnings — is that broken?

No. Those @deepseek-ai/* packages resolve at runtime to the global dsh install rather than into the profile. The official plugin market prints the same warnings.

It failed to start with EADDRINUSE

Another dsh is still running — often one whose terminal window was closed, leaving the process adopted by the system. Find it with lsof -nP -iTCP:3080 -sTCP:LISTEN and kill it.

I installed a plugin and nothing changed

The profile is read at boot, so a newly installed plugin needs a restart. Refreshing the page is not enough.

How are credits counted?

Semantic Scholar, PubMed and Google Scholar searches, auto-cite and figures are billed per call, and every answer ends with what it spent and what is left. arXiv, bioRxiv/medRxiv and DOI full-text extraction are free.

Can I install just one of the packages?

Yes — the five are independent layers. Install dsh-ai4scholar for search alone, or dsh-slides for decks alone.