Prompt Archaeology: A Five-Stage Protocol for Documenting and Analysing Human-AI Research Dialogues

Abstract

Prompt Archaeology proposes a five-stage protocol for retrospectively documenting, classifying, and descriptively analysing the recoverable prompt exchanges that shaped a human-AI research project, including human-to-LLM, LLM-to-sub-agent, and cross-model prompts. As large language models become research instruments, records of questions, corrections, and appraisals can help reconstruct an inquiry. Prompt Archaeology treats these recoverable exchanges as retrospective research artefacts. The proposal defines five intended plain-text artefacts, an eight-type prompt taxonomy, a minimum reporting-and-redaction set, and three proposed automation routes. Effort estimates are planning figures, not measured implementation performance. A partial, author-produced and author-coded pilot records 208 source-labelled entries (127 human, 81 LLM). These are recorded labels, not independently authenticated actor identities. The eight-type partition of the 127 human-labelled prompts is arithmetically consistent, but topic- and day-level aggregates remain unreconciled. A separately auditable Stage-1 inclusion/exclusion record is not established in the available pilot package. The earlier 30% topic-rate contrast is withdrawn; the pilot does not establish learning, knowledge quality, a causal effect, or a validated cross-model dispatch. (i) Contribution: a proposed five-stage documentation workflow, an eight-class taxonomy, explicit reporting contracts, and proposed automation routes; (ii) Evidential status: a partial, self-coded illustration; no complete audited implementation, public dialogue dataset, or external empirical validation is supplied; (iii) Evidence boundaries: retained-hash agreement tests internal consistency, not external origin, signatures, authorisation, enforcement, or legal compliance. External human inter-coder reliability and repeated-run stability remain open; an LLM-to-LLM agreement diagnostic is not human inter-rater validation; (iv) Positioning: relates to protocol analysis, dialogue-act tagging, qualitative reporting, and agent provenance literature. Surveys and other systems' security mechanisms motivate the proposal without validating its schema, tooling, or protections. This corrective v1.1 record remains an advanced methods research draft. It publishes the synchronised English and German manuscript PDFs only. Private dialogue material, CaseStudy2, implementation fixtures, and draft datasets are outside this record. Further case studies, corpus reconciliation, and independent human coding remain future work. Changes in Version 1.1 (October 2026) Corrective revision following the September inventory; no new empirical validation is claimed. Reframed the protocol, artefacts, automation routes, and pilot as an unvalidated proposal and partial illustration. Qualified recorded actor labels and the missing auditable Stage-1 record; withdrew the unsupported 30% topic-rate contrast and bounded unreconciled aggregates. Separated internal hash consistency from external provenance, authentication, enforcement, and compliance. Corrected the scope of agent-provenance literature and distinguished LLM-to-LLM agreement from independent human inter-coder reliability. Updated candid AI assistance disclosure and publication availability. DE and EN contain the same corrections, 34 cited keys, and eight labels. Released two separately readable manuscript PDFs. The historical v1.0 combined PDF remains with v1.0; no new combined PDF, private dataset, or Paper-2 manuscript is uploaded. Changes in Version 1.0 (June 2026) Initial release. Five-stage protocol with defined cumulative artefacts and an eight-type prompt taxonomy. Minimum Reporting and Redaction Set, plus three automation routes with effort estimates. Single illustrative pilot (208 prompts) with a full, mutually exclusive type distribution; presented as illustration, not validation. Prominent limitations: circular self-coded demonstration, recommended-but-uncomputed reliability metrics, single case study. DE/EN: English and German PDFs published synchronously; combined PDF generated from EN plus DE.