Concept Index

Purpose: A term-to-documents map for the workshop materials. Covers ~135 key concepts grouped by theme. Each entry names the concept, gives a one-line description, and lists the documents where it is treated substantively. Primary document listed first.

Intended uses: (1) quick navigation — find where a concept is covered; (2) data source for the GitHub Pages network visualisation and wiki-style hover links.

Tools and modes

Claude Desktop — the chat-based interface; Projects, custom instructions, file upload per session → A1.desktop-vs-code, A2.desktop-basics, A.concept.under-the-hood, A12.examples-claude-desktop-research

Claude Code — the terminal-based agent; reads files directly, runs commands, maintains project context → A1.desktop-vs-code, A3.code-basics-non-programmers, A.concept.under-the-hood

Claude Cowork — the browser-based collaborative interface; positioned between Desktop and Code → A1.desktop-vs-code

harness — the software layer that wraps an LLM and gives it tools, file access, memory, and the ability to act; separates the model (what reasons) from the operational environment (what it can do); Claude Code, Gemini CLI, Aider, and OpenCode are all harnesses → A4.conceptual-vocabulary, A.adjacent.tools

agent / agentic mode — Claude taking sequences of actions autonomously (read, write, run) rather than answering once → A.concept.agents, A.issue.plan-mode, A.issue.privacy-security

agent orchestration — Level 4 AI use: multi-step autonomous pipelines, API scripting, spawning subagents; the human role shifts from doing to specifying and reviewing → B.adoption-spectrum, A.concept.agents, C.leveling-packages, A11.examples-claude-code-researchers

plan mode — Shift+Tab; separates proposal from execution; Claude plans and waits for approval before acting → A.issue.plan-mode, A.concept.agents

context window — the total text Claude can hold at once; shapes what it knows and what it forgets → A.concept.under-the-hood, A.token.management, A.issue.context-scope

CLAUDE.md — the project configuration file Claude reads automatically at session start → A9.markdown-project-memory, A.issue.personalisation, A.issue.context-scope, B.trust

memory (Claude Code) — Claude-written persistence at ~/.claude/MEMORY.md; distinct from user-written CLAUDE.md → A.concept.memory, A.issue.personalisation

skills / slash commands — reusable task prompts invoked as /command-name; global or project-local → A14.skills-for-researchers, A.skills-ecosystem, A.issue.personalisation, A.issue.nodejs-basics

hooks — shell commands that run automatically on Claude Code events (PostToolUse, Stop, etc.) → A.issue.personalisation, A.issue.plan-mode

MCP (Model Context Protocol) — the standard for connecting external tools to Claude (Zotero, Slack, databases) → A4.conceptual-vocabulary, A.issue.personalisation, A.setup.settings-local, A.skills-ecosystem

RAG (Retrieval-Augmented Generation) — retrieve relevant documents, then generate an answer grounded in what was retrieved; the default pattern behind most AI-assisted research, whether or not the tool names it; fails structurally when the needed context was never written down → A4.conceptual-vocabulary, B.knowledge-curation

upload dance — the repeated manual file upload friction in Claude Desktop; resolved by Claude Code's direct file access → A.issue.upload-dance, A1.desktop-vs-code

installation (Claude Code on Windows) — step-by-step Windows native setup for non-programmers: Git, Node.js LTS, Claude Code via npm, Obsidian; winget primary with direct-installer fallback; includes troubleshooting → D.tutorial.setup, A3.code-basics-non-programmers, A.setup.windows-wsl, A.issue.nodejs-basics

terminal basics — the four PowerShell commands a non-programmer needs to use Claude Code: cd, dir, mkdir, and up-arrow for history → D.tutorial.setup, A3.code-basics-non-programmers

tutorial sequence — the three-step guided path from installation to sustained habit: setup → first project → first month → D.tutorials, D.tutorial.setup, D.tutorial.firstproject, D.tutorial.first-month

token management — staying within context limits; what to include, when to start fresh → A.token.management, A.issue.context-scope, A.issue.costs-quotas

settings.local.json — per-project local permissions file; controls which tools Claude can use → A.setup.settings-local, A.issue.team-claude

model selection — when to use Sonnet vs. Opus vs. Haiku; task-type calibration → A.issue.model-selection, A.issue.costs-quotas

costs and quotas — Claude Pro quota mechanics, rate limits, and API vs. subscription economics; when moving to the API becomes relevant for researchers → A.issue.costs-quotas, A.issue.model-selection, A.token.management

image analysis — using Claude as a visual reader for manuscript photographs, maps, and archival diagrams; what works (description, readable-text transcription, table extraction, comparison) and what fails (degraded script, non-Latin hands, spatial measurement) → A.working-with-images, A7.working-with-pdfs

pip / Python setup — Python's package installer; encountered when following setup instructions for MCP servers and tools; uv is the recommended modern alternative with fewer path conflicts → A.issue.python-basics, D.tutorial.setup

npm / Node.js / npx — the JavaScript runtime and package ecosystem; no longer required for Claude Code installation itself but still needed for npx skills add and npm-based MCP servers → A.issue.nodejs-basics, A.issue.personalisation

Project infrastructure and personalisation

context engineering — deliberately designing the persistent information architecture Claude operates within across sessions, rather than crafting individual prompts; the discipline that defines the L1→L2 transition on the adoption spectrum → A4.conceptual-vocabulary, C.leveling-packages, B.adoption-spectrum

prompt engineering — optimising individual prompts for better single-request responses; contrasted with context engineering, which builds the persistent foundation every request depends on → A4.conceptual-vocabulary, C.leveling-packages

project memory — using CLAUDE.md, _history.md, and progress files to maintain context across sessions → A9.markdown-project-memory, A.markdown.history, A.issue.personalisation, A.issue.research-data-memory

personalisation system — two-layer model: Claude Code mechanisms (CLAUDE.md, Memory, Skills, Hooks, MCP) plus project accountability files → A.issue.personalisation, A.issue.team-claude

meta-documents — _history.md, _ownership.md, progress.md, _index.md as a system: making a project legible to Claude and to yourself across sessions → A.markdown.meta-docs, A.issue.personalisation

_history.md — the project narrative: decisions, pivots, reasoning — not a task list → A.markdown.history, A.markdown.meta-docs, A.issue.personalisation

_ownership.md — the attribution record: collaboration mode per document, intellectual contributions that are unambiguously yours, ready-to-use disclosure statement → A.markdown.meta-docs, B.ownership

project legibility — the quality of a project being understandable to a returning Claude session or new collaborator without re-explanation → A.markdown.meta-docs, A9.markdown-project-memory, A.markdown.history

session-boundary habits — what to ask Claude at session start (read context, summarise state) and session end (append progress, record decisions) → A.markdown.meta-docs, A5.shortcuts-daily-usage

team configuration — sharing Claude setup via git; shared CLAUDE.md and skills as collective epistemic infrastructure → A.issue.team-claude, B.trust

global vs. local — ~/.claude/ (always active) vs. project .claude/ (active when launched from project) → A.concept.global-vs-local, A.issue.personalisation

bounding — defining clear edges for a task without predetermining the answer → A.issue.bounding, A.issue.context-scope, C.rules

form as bound — using official external structures (grant forms, submission guidelines) as prompting context so the form itself enforces the constraints; prevents plausible-but-non-compliant output → A.issue.bounding, A.issue.context-scope

prompt debugging — diagnosing failing output by root cause: task misunderstood, wrong context, or underbounded; reading response signals; the iterate-or-restart decision rule → A.issue.prompt-iteration, A.issue.bounding, A.issue.context-scope

project setup (from scratch) — creating a new research project with git init, start.md, optional CLAUDE.md, and an initial PDF workflow; the concrete path to Level 2 adoption → D.tutorial.firstproject, C.leveling-packages, A9.markdown-project-memory

multi-project planning — managing context across several active projects simultaneously; three approaches: parent-directory launch, explicit per-session loading, and a dedicated planning project → A.multi-project-planning, A9.markdown-project-memory

framing effect — what you include in a prompt shapes what Claude thinks the task is, not just what it knows → A.issue.context-scope, B.trust, A.issue.bounding

verification checkpoint — a defined stopping point to check output before proceeding; verification as architecture not afterthought → B.trust, A.critical.limitations, C.leveling-packages, C.rules

domain knowledge layer — the middle tier of project memory: _logic/ or named spec files (data definitions, analysis plans, transformation logic) read explicitly per session type, not loaded automatically; bridges CLAUDE.md routing and on-demand meta-documents → A.issue.research-data-memory, A.issue.personalisation, A9.markdown-project-memory

intermediary documents — themes.md, positions.md, gaps.md created incrementally during large-corpus work to preserve the researcher's evolving synthesis across sessions; the cognitive exoskeleton that makes a multi-week field synthesis project navigable → C.workflows, A.issue.research-data-memory, A.issue.research-wiki

research wiki / LLM wiki pattern — a living knowledge base built incrementally in a project folder: raw sources layer, wiki synthesis layer, CLAUDE.md as schema; three operations (Ingest a source, Query with full context, Lint for internal consistency); scales project memory from per-session notes to a navigable corpus → A.issue.research-wiki, A.issue.research-data-memory, C.llm-as-reader

Critical limitations and output risks

hallucination — Claude invents plausible-sounding facts, citations, and statistics confidently → A.critical.limitations, C.dangers, B.lifecycle.2.literature, A.working-with-images

vibe research — the entire argument or literature review feels correct but is not grounded in actual sources → A.critical.limitations, C.dangers, C.llm-as-reader, C.resources

execution vs. interpretation — the core distinction: Claude is reliable for rule-following tasks, unreliable for generating arguments and judgments → A.critical.limitations, B.lifecycle.4.dataanalysis, B.trust

scope misalignment — Claude answers a question adjacent to the one you asked, not the one you meant → B.trust, C.dangers

qualitative coding — AI-assisted coding of qualitative data introduces reliability and validity complications → B.lifecycle.4.dataanalysis, C.dangers, C.resources

false systematicity — repeatable AI-assisted workflows create an impression of rigour that the underlying outputs may not warrant; errors replicate silently at scale rather than staying visible in single sessions → C.leveling-packages, B.adoption-spectrum, B.lifecycle.0.iteration, A.issue.research-wiki

knowledge base error compounding — in an AI-maintained wiki, a hallucinated or misread fact propagates silently through cross-references; the lint operation checks internal consistency but not source-vs-wiki accuracy; structurally isolated claims (no inbound cross-links) are the highest-risk entries → A.issue.research-wiki, C.llm-as-reader

overconfidence — models use higher-confidence language when generating incorrect information (34% effect) → A.critical.limitations

Epistemic risks

sycophancy — Claude's structural tendency to produce outputs that align with what you want to hear; a consequence of RLHF, not deception → B.epistemics, A.critical.limitations, C.dangers, B.trust

confirmation loop — sycophancy at project timescale: hypothesis in CLAUDE.md → session confirms it → stronger framing next session → B.epistemics, C.dangers, B.autonomy

tunnel effect — Claude operates within the frame you provide; cannot say "your question is wrong" → B.epistemics, C.dangers

team groupthink — shared CLAUDE.md + shared hypothesis = same sycophantic tendency running in parallel for each team member → B.epistemics, B.team-ai, C.dangers

involuntary centroid capture — AI's gravitational pull toward the statistical centre of published literature; silently replaces peripheral intuitions; in skills: without intervention Claude defaults to the statistical average response → B.centroid-periphery, B.lifecycle.1.creativity, C.dangers, A.skills-ecosystem

tractability drift — peripheral research questions migrate toward more central, AI-tractable formulations without a visible decision point → B.centroid-periphery, B.autonomy

center vs. periphery — center work (synthesis, standard methods) is legitimate and AI-excellent; periphery (novel questions, gaps) requires a different AI relationship → B.centroid-periphery, B.lifecycle.1.creativity

adversarial prompting — running a fresh session with an explicit brief to argue against your current position → B.epistemics, A15.agent-personalities

cross-model checking — running the same question through Claude and Gemini; divergence is diagnostic, not a winner/loser determination → B.epistemics, A.adjacent.gemini

detectability spectrum — epistemic risks (confirmation loops, groupthink) are less detectable but more damaging than output risks (hallucination) → C.dangers

epistemic opacity — when an AI-built knowledge base's weighting function becomes invisible; researchers interact with the tool's model of a corpus rather than their own; a structural risk at the "Claude reads for you" end of the spectrum → C.llm-as-reader, B.epistemics

Autonomy and long-term risks

autonomy — remaining the person who decides the question, evaluates the answer, and takes responsibility; AI can expand or erode it → B.autonomy, B.experience

epistemic capture — Claude reframes your question and you follow the new frame without noticing the shift → B.autonomy, B.lifecycle.1.creativity

atrophied judgment — the ability to evaluate AI output erodes from not exercising it; self-concealing because confidence does not decline with capability → B.autonomy, C.dangers, C.llm-as-reader

agenda drift — research questions migrate toward what AI handles well; away from the anomalous and genuinely unknown → B.autonomy, C.dangers

voice homogenisation — sustained exposure to Claude's output patterns gradually shapes your own writing style → B.autonomy, C.dangers

peer substitution — consulting Claude instead of colleagues; Claude cannot provide genuine intellectual disagreement and shared stakes → B.autonomy, C.dangers

Socratic mode — using Claude to stress-test your idea rather than generate one; the appropriate AI role at the research frontier → B.lifecycle.1.creativity, B.centroid-periphery, A15.agent-personalities

Ownership and attribution

ownership — who contributed what: ideas, structure, evidence, language, evaluative judgment → B.ownership, B.experience

attribution — crediting AI's role accurately in publications, reports, and submissions → B.ownership, E.journal-funder-policies, E.muni-guidelines

disclosure — when, how, and at what level to declare AI assistance in academic work → B.ownership, E.muni-guidelines, E.cuni-researcher-recommendations, E.journal-funder-policies

third-party problem — AI entering collaborative relationships as a silent participant colleagues cannot question or calibrate against → B.team-ai, C.dangers

epistemological fragility — code or analysis you cannot explain is analysis you cannot defend at peer review → A.critical.limitations, B.ownership

Research lifecycle

research question formation — the most interpretive phase; Claude's centroid picture misleads; use it Socratically → B.lifecycle.1.creativity, B.lifecycle.md, B.centroid-periphery

literature finding — structurally fragile: training cutoff, no paywall access, hallucination risk; distinct from synthesizing → B.lifecycle.2.literature, B.lifecycle.md, A.literature-discovery

literature synthesis — Claude is genuinely productive here once you have the sources; map themes, tensions, connections → B.lifecycle.2.literature, B.lifecycle.md

data collection — mostly human; AI supports preparation and post-collection processing; the micro-discovery loop is the key contribution → B.lifecycle.3.datacapture, B.lifecycle.md, A.issue.research-data-memory

data analysis — execution (AI-excellent) vs. interpretation (human); qualitative coding has specific complications → B.lifecycle.4.dataanalysis, B.lifecycle.md, A.issue.research-data-memory

manuscript writing — mechanics are AI-assisted; argument-building is interpretive; writing as thinking → B.lifecycle.5.manuscript, B.lifecycle.md

micro-discovery loop — processing incoming material close to collection so findings redirect further collection early → B.lifecycle.0.iteration, B.lifecycle.3.datacapture

PDF-to-markdown workflow — converting PDF files to editable markdown so Claude can read, search, and cite them efficiently within a project folder → D.tutorial.firstproject, A7.working-with-pdfs

iterative research / fail quickly — discovering problems early when correction is cheap; AI tightens the feedback loops → B.lifecycle.0.iteration, B.lifecycle.md, C.start-here, C.rules

AI co-scientist — paradigm for AI as a systematic research partner: hypothesis generation, expert-level critique, and iterative refinement cycles; designed around the scientist-in-the-loop principle even at maximum autonomy → A11.examples-claude-code-researchers, B.lifecycle.1.creativity, C.calibration, A.concept.agents

transformation log — step-by-step record of every data transformation (input → operation → output → rationale) with a canonical-version header at the top; the most underrated file in data-heavy research for reproducibility and debugging → A.issue.research-data-memory, B.lifecycle.3.datacapture

data variants — a named-variants file documenting each dataset version with purpose, path, creation date, and a valid/confounded decision rule; prevents the "data_final_v2_REAL_use_this_one.csv" syndrome → A.issue.research-data-memory, B.lifecycle.3.datacapture

reader spectrum — the range from "Claude assists your reading" (you engage with sources; Claude organises) to "Claude reads for you" (ingestion pipeline; knowledge base grows faster than you read); both ends are legitimate; the mistake is arriving at either end by accident → C.llm-as-reader, A.issue.research-wiki, B.autonomy

syntopical reading — reading across an entire field simultaneously (Adler); interrogating many authors on a shared question rather than reading one source at a time; requires multi-session infrastructure and intermediary documents → C.workflows, B.lifecycle.2.literature, C.llm-as-reader

two fragile moments — question formation and literature finding are dangerous in kind: AI actively misleads, the misleading is hard to detect, damage compounds downstream → B.lifecycle.md, C.dangers

Team and collaboration

team AI use — when multiple researchers use AI on shared work; the third-party problem, adoption asymmetry, disclosure norms → B.team-ai, A.issue.team-claude

adoption asymmetry — uneven AI use across a team creates invisible capability gaps and trust imbalances → B.team-ai

framing rotation — deliberately assigning different starting hypotheses to different team members to structurally build in disagreement → B.epistemics, B.team-ai

Slack as worktable — integrating Slack conversations into Claude's context; copy-paste workflow vs. MCP server → A.issue.slack-workspace, B.team-ai

Privacy and security

data privacy — Desktop sends data to Anthropic servers; Code runs locally; matters for interviews, unpublished data, GDPR-covered material → A.issue.privacy-security, C.dangers

GDPR — EU data protection implications for using cloud AI with research data → A.issue.privacy-security

agentic risk — Claude Code in agentic mode can read, write, delete files; mistakes may be irreversible → A.issue.privacy-security, C.dangers, A.concept.agents

local vs. cloud processing — the key privacy distinction: Claude Code (local, data stays on machine) vs. Desktop (cloud, Anthropic servers); "local" does not mean "private" — prompts still travel to Anthropic servers → A.issue.privacy-security, A1.desktop-vs-code, A.concept.under-the-hood

.claudeignore — a file placed in the project root listing files and folders Claude should never read; independent from .gitignore; protects credentials, raw personal data, and embargoed material in agentic sessions → A.issue.privacy-security, A.setup.settings-local

Integrations and adjacent tools

Zotero — connecting Claude to your reference library; reading, searching, suggesting connections → A10.claude-and-zotero, A.setup.zotero-mcp, A.literature-discovery

NotebookLM — Google's document-grounded AI; podcast generation, study guides; complements Claude for passive orientation → A.adjacent.tools, B.TH.personal.use-cases

Gemini — Google's model; primary cross-model adversarial partner; Gemini Deep Research for literature orientation; Gemini CLI is Google's harness equivalent of Claude Code → A.adjacent.gemini, B.epistemics, A.literature-discovery, A.adjacent.tools

Aider / OpenCode — open-source model-agnostic coding harnesses; Aider (aider.chat) connects to Claude, GPT-4, Gemini, or local models; OpenCode (opencode.ai) supports 75+ models with a privacy-first architecture → A.adjacent.tools

Obsidian / Logseq — personal knowledge management tools with Claude Code integration → A.obsidian-logseq

Graphify — graph-first alternative to the research-wiki pattern; deterministic tree-sitter parsing for code, LLM extraction for prose/PDFs/images, Leiden clustering; savings come from querying an already-built graph, not from cheaper ingestion → A.issue.research-wiki

git — version control as safety net for Claude Code sessions; reversibility for agentic work → A.git-for-researchers, A.issue.team-claude

Excalidraw — diagramming tool; project-local skill for creating visual arguments → research-lifecycle.excalidraw

WSL — Windows Subsystem for Linux; required for Claude Code on Windows → A.setup.windows-wsl

Institutional and regulatory

MUNI guidelines — Masaryk University AI recommendations (non-binding, 2023); no researcher/publication policy yet → E.muni-guidelines

CUNI researcher recommendations — Charles University researcher guidance (freely distributable, co-authored by MUNI); best available Czech researcher framework → E.cuni-researcher-recommendations

journal and funder policies — publisher and funder AI disclosure requirements; convergence on five shared rules; Horizon Europe mandatory disclosure → E.journal-funder-policies

academic integrity — the "own the result" standard: can you explain it, defend it, stand behind it? → E.muni-guidelines, B.ownership

Experience and practice

experience (emotional dimension) — control/ownership axis; joy and curiosity vs. overwhelm and FOMO; real vs. anxiety-driven productivity → B.experience

trust (designed relationship) — trust is a property of the researcher-tool relationship, not of the tool itself; built through scope control, verification, and project infrastructure → B.trust, A.issue.personalisation

active reading — reading through iterative questions rather than asking for summaries; the most transferable pattern across use cases → B.TH.personal.use-cases, A.critical.limitations, D.tutorial.firstproject

reading note — a structured synthesis document (per-source arguments, shared themes, open questions, next steps) produced by Claude from a source set; editing it is the ownership and verification practice → D.tutorial.firstproject, A.workflows.notes-reading-writing

agent personalities — Devil's Advocate, Critical Reviewer, Socratic Dialogue Partner — configuring Claude's role explicitly → A15.agent-personalities, B.epistemics

personal use cases — nine first-person cases from the presenter: dissertation review, course prep, data analysis, software building, project management, this workshop → B.TH.personal.use-cases

institutional use cases — six cases where AI is commissioned for group/department tasks: website audit, curriculum guide, thesis archive, data snapshots, synthesis, research mapping → B.TH.institutional.use-cases

periphery task problem — valuable institutional tasks that stay permanently undone because they are never urgent enough to justify the sustained effort they require → B.TH.institutional.use-cases

institutional commission — using AI for research group or department tasks rather than personal assistance; changes the threshold for what is worth starting → B.TH.institutional.use-cases, B.usecases

false closure — having a finished output without anyone owning it; the output becomes a substitute for the work rather than its beginning → B.TH.institutional.use-cases, B.ownership

maintenance debt — periphery tasks done at AI-assisted scale produce artefacts (guides, visualisations, reports) that need owners and ongoing updates that nobody budgeted for → B.TH.institutional.use-cases

adoption spectrum — five-level continuum of AI-assisted research integration (occasional querier → regular conversation partner → context-aware collaborator → workflow integrator → agent orchestrator); position reflects structural integration depth, not technical skill → B.adoption-spectrum, C.leveling-packages, C.tasks, C.why

leveling (transition guide) — what to conceptually grasp, what to try first, and what practice to sustain when moving from one adoption level to the next → C.leveling-packages, B.adoption-spectrum

post-workshop adoption — week-by-week first-month guide for making Claude Code a real habit: five abandonment patterns and their diagnostic signals; where adoption sticks or fades → D.tutorial.first-month, C.leveling-packages, B.adoption-spectrum, A16.personal-life-management

capacity ceiling — Claude expands what is achievable within fixed time; it does not automatically compress when tasks happen or save clock hours; the freed capacity is a decision the researcher makes → B.time, C.why, B.experience

expansion trap — freed AI time fills immediately with new tasks unless deliberately protected; Parkinson's Law applied to AI-assisted work → B.time

ten simple rules — a condensed good-practice guide: setup and process rules (1–4), epistemic rules (5–8), sustainability rules (9–10); PLoS format with named failure mode per rule → C.rules, C.calibration

getting started / orientation — the five prerequisite concepts for self-directed readers: working directory, markdown as bridge, CLAUDE.md, fail quickly, adoption spectrum → C.start-here

prompt gallery — a collection of copy-paste-and-adapt prompts for common humanities research tasks, annotated with why each is phrased as it is → C.prompts, C.workflows

quick reference / cheatsheet — condensed daily-use reference for terminal commands, session starts, and starter prompts; the one-page resource for participants after the workshop → C.cheatsheet, D.handout

further reading — curated reading list for participants going deeper: Mollick on co-intelligence, Pepinsky on execution vs. interpretation, Hall on vibe research, digital humanities practitioners, peer-reviewed qualitative methods scholarship → C.resources