What you can do: tasks by adoption level

How to use this document. Find your current adoption level (see B.adoption-spectrum if unsure), then read what is available to you now. The list is cumulative — tasks at lower levels remain available at higher ones. Tasks at the next level are what you are working toward.

For each task: what AI does, what the human retains, and a pointer to the fuller treatment.

For tasks AI is not reliable for, see the closing section.

Level 1 — available in any web chat, no setup required

These tasks work with pasted text or a single uploaded file. No CLAUDE.md, no Claude Code, no prior configuration.

Concept or term explanation What AI does: explains a concept, term, or field orientation in plain language, at the level you specify. Human retains: judgment about whether the explanation is accurate for your domain; verification of any specific claims. → A4.conceptual-vocabulary, A.concept.under-the-hood

Paper summary (pasted or uploaded) What AI does: extracts the main argument, methods, evidence, and conclusions from a paper you paste or upload. Human retains: assessment of whether the summary captures what matters for your purposes; verification of any specific claims. → B.usecases, A7.working-with-pdfs

Research question brainstorming What AI does: generates multiple framings, angles, and sub-questions from a broad topic you describe. Human retains: all decisions about what is interesting, novel, or feasible; beware centroid pull toward the already-known. → B.lifecycle.1.creativity, B.centroid-periphery

Socratic refinement of a research question What AI does: asks clarifying questions, reflects back what it heard, surfaces assumptions, proposes more specific framings — without deciding which is better. Human retains: every decision; Claude facilitates, the researcher chooses. → B.usecases, B.lifecycle.1.creativity

First draft of a short text What AI does: drafts an abstract, introduction paragraph, email, cover letter, or short summary from your bullet points or notes. Human retains: all content decisions; the draft is a starting point, not a result. → B.lifecycle.5.manuscript, B.usecases

Style and writing check What AI does: flags long sentences, passive voice, register inconsistencies, structural weaknesses; optionally proposes rewrites. Human retains: all decisions about what to change; mechanical flags are reliable, structural feedback requires evaluation. → B.usecases, A.issue.bounding

Translation and language support What AI does: translates passages (Latin, German, French, Czech, and others); explains idioms, abbreviations, archaic usage; helps read unfamiliar scripts or registers. Human retains: domain-specific interpretation; verification of specialist terms; all philological judgment. → A13.examples-dissinet-usecases

Argument structuring What AI does: takes your bullet points and proposes a logical sequence; identifies what is missing or redundant; proposes a section structure. Human retains: all content; structure is a proposal, not a prescription. → A.issue.bounding, B.lifecycle.5.manuscript

"What do you know about X?" What AI does: gives an orientation to a topic, names key figures and debates, sketches the state of a field — at a conversational level. Human retains: verification of everything cited; this is orientation, not a literature review; hallucination risk on specific sources. → A.critical.limitations, B.lifecycle.2.literature

Level 2 — requires persistent project context

Set up a CLAUDE.md file (Claude Code) or a Project description (Claude.ai). One-time investment, compounding return. See C.leveling-packages for the 30-minute setup.

All Level 1 tasks remain available — now project-aware.

Project-aware paper assessment What AI does: reads a paper and assesses its relevance to your specific project, using your project context to judge what matters and what doesn't. Human retains: all final relevance judgments; Claude's assessment is informed but not authoritative. → B.usecases, A9.markdown-project-memory

Draft feedback knowing your project What AI does: reads your draft and gives feedback against your stated goals, audience, and standards — not generic "good writing" criteria. Human retains: all decisions about what to change; argument-level feedback needs evaluation. → B.usecases, B.lifecycle.5.manuscript

Dissertation or thesis review What AI does: reads a full dissertation and produces: core argument summary, chapter-by-chapter critique, strongest/weakest points, suggested viva questions. Human retains: all substantive judgments; the viva questions in particular need triage — Claude produces generic and excellent questions together. → B.usecases, B.TH.personal.use-cases

Grant section drafting What AI does: drafts a specified section (State of the Art, Impact, Methodology) within character limits, mirroring call language, staying inside the researcher's stated claims. Human retains: all factual claims; verification of character count; judgment about whether the framing is defensible; all novelty claims. → B.usecases, B.lifecycle.5.manuscript

Source cross-referencing against project notes What AI does: takes a new source and maps it against your project's key claims, existing notes, and open questions — identifying where it confirms, contradicts, or adds to what you already have. Human retains: all interpretive judgments about what the mapping means. → A9.markdown-project-memory, B.lifecycle.2.literature

Gap identification in your Zotero collection What AI does: searches your Zotero library for materials on a topic; then proposes what key authors or works it would expect to find that are not in your library. Human retains: verification of all proposed gaps — this is Claude's best guess from training data, not authoritative bibliography; recent publications and specialist literature may be missed or invented. → B.usecases, A10.claude-and-zotero

CLAUDE.md creation for a new project What AI does: interviews you with 5–7 questions, then drafts a project context file that will inform every future session. Human retains: verification that the CLAUDE.md accurately reflects the project; updating it as the project evolves. → B.usecases, A9.markdown-project-memory

Extended research design dialogue What AI does: engages in multi-round dialogue on research design, argument development, hypothesis testing — building on earlier turns in the same session. Human retains: every decision; the dialogue is a thinking tool, not a decision-making process. → B.usecases, B.lifecycle.1.creativity, B.epistemics

Level 3 — requires Claude Code + workflow design

Claude Code runs in your project folder. It can read, write, and process your actual files. See A3.code-basics-non-programmers for getting started.

All Level 1 and 2 tasks remain available — now at file and batch scale.

Batch PDF triage What AI does: works through a folder of PDFs (archive scans, downloaded papers, court documents), assessing readability, content type, names, dates, and structural features; writes results to a table file. Human retains: all triage decisions; readability assessments are a starting point; Claude's reading of manuscript images is imperfect. → B.usecases, A7.working-with-pdfs

Corpus extraction with a consistent schema What AI does: applies a defined extraction schema (fields, categories, formatting rules) consistently across every document in a folder. Human retains: schema design; verification of a sample of outputs; judgment about ambiguous cases. → A7.working-with-pdfs, B.lifecycle.3.datacapture

Structured bibliography / research log maintenance What AI does: adds new entries to a running structured file (bibliography, reading notes, source log) in a consistent format; de-duplicates; flags inconsistencies. Human retains: content accuracy; judgment about what to include. → A9.markdown-project-memory, A10.claude-and-zotero

Interview transcript coding What AI does: applies a defined coding scheme to interview transcripts; flags ambiguous passages that require human review; produces a coding table. Human retains: scheme design; all ambiguous cases; intercoder reliability verification; all interpretive conclusions. → B.lifecycle.4.dataanalysis, A.critical.limitations

Project folder reorganisation What AI does: reads the current folder, proposes a sensible structure, lists every existing file with its suggested destination, optionally moves the files. Human retains: approval of the proposed structure; all go/no-go decisions before files are moved. → B.usecases, A.concept.global-vs-local

Literature synthesis from a document set What AI does: reads a folder of papers (or a Zotero export) and produces a structured synthesis: key themes, points of consensus and disagreement, notable gaps — organised by the questions you specify. Human retains: verification against the sources; all interpretive conclusions; the synthesis is a scaffold, not a final product. → B.lifecycle.2.literature, A.literature-discovery

Multi-format output generation What AI does: takes a single source (interview transcript, structured data file, research notes) and generates multiple output formats — structured table, narrative summary, annotated bibliography entry, data snapshot — in one session. Human retains: all content decisions; verification of each output format. → A8.working-with-docx-xlsx, A6.zero-coding-workflows

Consistent feedback across a student cohort What AI does: applies a defined feedback template to multiple student papers, producing structured feedback for each in a consistent format. Human retains: all substantive feedback decisions; triage of which AI-generated comments to keep; final feedback delivery. → B.TH.institutional.use-cases, B.TH.personal.use-cases

Systematic source annotation What AI does: adds structured metadata to a document set — time period, source type, named entities, themes — as a first-pass annotation layer. Human retains: verification of every annotation; corrections; interpretive categorisation. → B.lifecycle.3.datacapture, A7.working-with-pdfs

Level 4 — requires programming skill or close collaboration

Multi-step automated pipelines, API scripting, subagent orchestration. See B.adoption-spectrum for an honest assessment of whether this level is right for your situation.

All Level 1–3 tasks remain available, now as components in larger pipelines.

Automated document ingestion pipeline What AI does: scans → extracts → classifies → stores incoming documents to a structured database, without session-by-session human initiation. Human retains: pipeline design; verification gates at each step; exception handling; all schema and classification decisions. → B.adoption-spectrum, B.TH.institutional.use-cases

Multi-agent research synthesis What AI does: different subagents process different document sets or sub-questions; results are merged and synthesised in a final step. Human retains: task decomposition and agent briefing; review of each agent's outputs; synthesis judgment; all interpretation. → B.adoption-spectrum, A.concept.agents

Scheduled processing of incoming materials What AI does: processes newly arrived documents (emails, downloaded papers, data exports) on a schedule without manual initiation each time. Human retains: pipeline design; review cadence; exception handling; all decisions triggered by what the pipeline finds. → B.adoption-spectrum

Template-driven document generation from structured data What AI does: generates documents (reports, data snapshots, formatted outputs) from a database or spreadsheet, populated via API. Human retains: template design; data quality; review of generated outputs; all content decisions embedded in the template. → B.TH.institutional.use-cases, A8.working-with-docx-xlsx

Cross-project synthesis What AI does: reads across multiple research project directories and identifies connections, overlapping themes, potential synergies, or gaps. Human retains: all interpretive conclusions; verification of cross-project claims; research direction decisions. → B.TH.personal.use-cases, B.TH.institutional.use-cases

Tasks AI is not reliable for

These tasks are not on the list above because AI performance on them is systematically poor, unpredictable, or dangerous enough to require a direct warning.

Finding literature reliably — Claude will hallucinate citations, miss recent publications, and miss paywalled or specialist-field sources structurally. Use Zotero + Gemini Deep Research for discovery; use Claude only for synthesis once you have verified sources. → A.literature-discovery, B.lifecycle.2.literature, A.critical.limitations

Forming original research questions — Claude's training represents the statistical centre of published literature. Using it to generate research questions pulls toward the already-known; your peripheral intuitions are the ones worth protecting. Use Claude to stress-test a question you formed, not to form it. → B.lifecycle.1.creativity, B.centroid-periphery

Final interpretive judgment in qualitative analysis — pattern identification, coding, and extraction Claude can do at scale; deciding what the patterns mean, whether the interpretation is defensible, and what they imply for your argument is not delegatable. → B.lifecycle.4.dataanalysis, A.critical.limitations

Reliable translation of specialist historical documents — Claude translates well in common languages and registers; it is unreliable with technical historical vocabulary, heavily abbreviated Latin, damaged text, and non-standard orthography. Use it for orientation, not for final philological judgment. → A.critical.limitations