Promptotyping is an iterative, knowledge-driven working practice through which a project builds its own digital research artefacts out of structured research data and the project knowledge it maintains, using Context Engineering and Agentic Engineering and organised around four recurrent forms of work (Preparation, Exploration, Distillation, Implementation). In the canonical repository convention, a knowledge/ folder holds versioned Knowledge Documents for requirements, data, decisions, and domain knowledge. This curated knowledge artefact guides implementation, persists across sessions, and receives consequential findings through write-back. The practice also permits another inspectable and revisable documentary environment. A Critical Expert in the Loop verifies LLM output at defined points.
This repository is the public Promptotyping portal. It holds the practice specification and its supporting material, the templates, examples, site and tools. The portal at https://dhcraft.org/Promptotyping/ integrates the specification, case evidence, tutorial, templates and canonical skill. It also serves the templates as stable, machine-readable addresses.
The four forms of work move from gathering raw material (Preparation) over probing what the data affords (Exploration) and distilling the findings into selective Markdown documents (Distillation) to iterative, verified development with an agentic coding tool (Implementation). Knowledge Document is the general term for a document of this kind, and it specialises into three types according to the kind of knowledge it holds, each with a diagnostic rule attached. Declarative Documents hold knowledge about the subject matter. Process Documents divide current handoff inputs, accepted future work, and the curated provenance index between handoff.md, plan.md, and journal.md. Action Documents, which the paper calls agent instruction documents, hold knowledge about how to act (CLAUDE.md, AGENTS.md); check these when output is formally wrong.
An overview of the practice is at dhcraft.org/Promptotyping/#overview.
The paper is maintained separately and will be published in the journal. The manuscript and the material bound to it, its specification, review tooling, figures and audits, are not part of this repository. On 14 October 2026 the author removed the manuscript from the public repository and website, and on 28 October 2026 the remaining manuscript-bound material followed. Earlier public availability remains part of the publication history.
This repository contains no third-party research data. Its material basis is recorded in knowledge/project.md. The cleaned German video transcripts under source-material/ and the sources of the Grounded Vault under vault/ are carried with their provenance, and benchmark tasks hold their own data contracts and provenance in the benchmark authoring area.
The portal is published with GitHub Pages directly from the repository root at https://dhcraft.org/Promptotyping/.
git clone https://github.com/DigitalHumanitiesCraft/Promptotyping.git
cd Promptotyping
python -m http.server 8000Open http://localhost:8000.
Two introductions to the practice (German): the explainer Promptotyping, Teil 1 covering the four forms of work, the documents, and Critical Expert in the Loop, and the live demo Teil 2, building a functional dashboard from 76 Excel files with Claude Code.
| Content | Location | On the site |
|---|---|---|
| Evidence layer of the practice (Grounded Vault) | vault/ | #vault |
| Promptotyping Bench (draft research programme) | Proposal, plan, task and tooling index, research map | No dedicated portal page |
| Knowledge Document templates | _content/promptotyping-document/ | #vorlagen |
| Convention (naming contract, frontmatter schema, structure) | _content/konvention.md | #konvention-v0.1 |
| Technology baseline for static research tools (draft) | _content/technology-baseline.md | — |
| Curated case studies | data/case-studies.json, depth pages in _content/case-studies/ | #use-cases |
| Guided first iteration | _content/tutorial.md | #tutorial |
| Extensions grown from practice | _content/praxis.md | #praxis |
| Canonical Promptotyping skill | DigitalHumanitiesCraft/promptotyping-skill (own repository) | #skills-promptotyping |
| Reusable system prompts (coding, writing) | _content/skills/ | #skills |
| Glossary | data/glossar.json, _content/glossar.md | #glossar |
| Durable site knowledge base (self-application) | knowledge/ | — |
| Live Process Inbox | knowledge/handoff.md | — |
| Source material, including video scripts | source-material/ | — |
| Dated reports and audits | snapshots/ | — |
The templates of the catalogue structure the documents of a Promptotyping knowledge base. Each is addressable under a stable latest URL of the pattern https://dhcraft.org/Promptotyping/promptotyping-document/{slug} and readable as raw Markdown under the machine address pattern below. A template applies only when its trigger holds. This Promptotyping portal records its material basis in knowledge/project.md; benchmark tasks hold their own data contracts and provenance in the separate authoring area.
| Slug | Function | Type |
|---|---|---|
| index | Navigation and glossary of the knowledge base | Declarative |
| project | Charter | Declarative |
| data | Material | Declarative |
| specification | Specification (requirements, decisions) | Declarative |
| user-stories | Specification (narrative scenarios) | Declarative |
| action-layer | Agent instructions (CLAUDE.md) |
Action |
| architecture | Architecture | Declarative |
| technology | Technology baseline for an artefact family | Action |
| domain-knowledge | Domain knowledge and reasoning layer | Declarative |
| design | Design | Declarative |
| testing | Quality assurance | Action |
| verification | Adversarial verification of claims | Declarative |
| journal | Provenance | Process |
| handoff | Process Inbox | Process |
| plan | Planning | Process |
| report | Reporting | Declarative (snapshot) |
| integration | Cross-project contracts | Declarative |
Template versions live in the mirrors' frontmatter and in data/promptotyping-documents.json, which is the catalogue this table follows.
The gallery at #use-cases curates publicly documented projects built with the practice; each card links the project's GitHub repository and, where available, live demo and video. The underlying data is data/case-studies.json, and the cases marked there with deep_page carry a depth page under _content/case-studies/.
Repositories that apply the practice declare which template a document follows in a template: frontmatter field:
template:
name: Vorlage Datengrundlage
version: <current template version, see data/promptotyping-documents.json>
url: https://dhcraft.org/Promptotyping/promptotyping-document/data
alias: https://dhcraft.org/Promptotyping/#promptotyping-document-dataThe url resolves in a browser via 404.html routing and requires JavaScript. For HTTP retrieval without JavaScript, the canonical machine address is the static Markdown file:
https://dhcraft.org/Promptotyping/_content/promptotyping-document/{slug}.md
.nojekyll in the repository root is required for GitHub Pages to publish _content/. Details in knowledge/specification.md (A4, A5, ADR-10).
index.html single-shell entry for the Promptotyping portal
404.html subpath-to-hash routing fallback
CLAUDE.md action layer of this repository (its own instance of the action-layer template)
.nojekyll required so GitHub Pages publishes _content/
knowledge/ knowledge base of the site itself
research-artefacts/ benchmark proposal and benchmark authoring area
source-material/ cleaned video transcripts
snapshots/ dated reports and audits
schema/ JSON Schema of the frontmatter core
tools/ consistency check, glossary build, tests
_content/ Markdown content (practice pages, templates, case studies, glossary)
assets/ CSS, JS, vendored libraries, fonts, logo
data/ JSON data (case studies, glossary, template table, vault index)
vault/ Grounded Vault instance, evidence layer of the practice
No framework, no build step; marked.js and js-yaml are vendored. The reasoning is documented in the technology baseline.
The site applies the practice it specifies. Its knowledge base in knowledge/ follows the templates the site publishes, each document carrying the template: field that points back to its own specification, and CLAUDE.md is the agent's entry point. The harness was Claude Code. The Co-Authored-By trailers of the commits name Claude without a model for the first phase from September to November 2025, and later Claude Sonnet 4.6, Claude Opus 4.7, Claude Opus 4.8, Claude Fable 5, Claude Opus 5 and Claude Opus 5.5; commits without a trailer do not record whether an agent was involved.
Checking is layered. Verification runs tools/check_consistency.py, which decides by rule where the site states the same thing in more than one place, together with the regression tests of the tools and the validator of the Grounded Vault (knowledge/verification.md). The practice's supporting claims are anchored in the Grounded Vault under vault/, which links every assertion to a source with provenance. Where agreement in substance is not decidable by rule, the check stays with the Critical Expert, and knowledge/verification.md records it as outstanding. The development history is in knowledge/journal.md.
- The Promptotyping Bench is a draft research programme. Its tooling tests supply engineering evidence and no study result (knowledge/benchmark-plan.md).
- The automatic checks decide consistency between the site's own statements. Whether the practice description is right in substance remains a matter of scholarly judgement.
- The templates other than
technologyare mirrors of the operator's vault and stay German.
This repository documents the Promptotyping working practice. The paper is maintained separately and will be published in the journal.
For the repository and site: Pollin, Christopher; Steiner, Christian: Promptotyping. Zenodo. https://doi.org/10.5281/zenodo.14160875 (concept DOI, always resolves to the current version)
Promptotyping originates in the author's dissertation and was first published in blog form:
- Pollin, Christopher: Modelling, Operationalising and Exploring Historical Information. Using Historical Financial Sources as an Example. Dissertation, University of Graz 2025. https://unipub.uni-graz.at/obvugrhs/12127700
- Pollin, Christopher: Promptotyping. Von der Idee zur Anwendung. DHCraft Blog, 24.04.2025. https://dhcraft.org/excellence/blog/Promptotyping/
- Pollin, Christopher: Promptotyping. Zwischen Vibe Coding, Vibe Research und Context Engineering. Blogpost, L.I.S.A. Wissenschaftsportal der Gerda Henkel Stiftung, 17.01.2026. https://lisa.gerda-henkel-stiftung.de/digitale_geschichte_pollin
Code (HTML, CSS, JavaScript, and the Python tooling under tools/, vault/tools/ and research-artefacts/promptotyping-bench/) is licensed under the MIT License, see LICENSE. Content (Markdown under _content/, knowledge/, the Grounded Vault under vault/, and the prose texts of the site) is licensed under CC BY 4.0. The split follows the carrier; source that runs is MIT, text that is read is CC BY 4.0. This repository contains no third-party research data.