# Chat Archive Index

Source: `new_experiments.zip -- ATWL workflow-design prompting-condition experiment`  
Generated: 2026-09-04 15:56 UTC  
Chats indexed: 38

## What this archive is

These 38 chats are almost all part of **one prompting-strategy experiment** about designing analytical
workflows and expressing them in **ATWL** (an "Artifact-Transform Workflow Language" for formalising
visual-analytics workflows), run across two LLMs (**Claude Sonnet 5** and **GPT-5.6-Terra**), two design
tasks (**VAST**: a Vastopolis flu-outbreak investigation; **VIS**: characterising IEEE VIS author
collaboration communities), and **8 prompting conditions** per model/task combination:

| Condition | What's supplied | What's asked for |
|---|---|---|
| `null` | task only | free-form plan, no ATWL |
| `ATWL` | ATWL spec + 1 example | prose + ATWL design |
| `ATWL+library` | spec + example + 53-workflow library *(pre-formalised in ATWL)* | design citing library workflows |
| `ATWL_2-stage` | → continues `ATWL+library` | + method-level enrichment from the cited source papers |
| `library_prose` | 53-workflow library *(natural-language summaries only)* | prose-only design citing workflows |
| `C6` | → continues `library_prose` | + faithful ATWL translation of the prose design |
| `C6+` | → continues `C6` | + self-critique of the translation, then a revised ATWL design |
| `C7` | → continues `library_prose` | + a **fresh** ATWL redesign (not a translation) |

That's 2 models × 2 tasks × 8 conditions = 32 chats (rows 1–32 below). The remaining **6 chats** are a
separate but related session: a researcher works with GPT-5.6-Terra to define a ~45-feature binary coding
scheme (motifs, structural indicators, model roles, analytical strategies, iteration targets) for ATWL
workflows, applies it to the 53-workflow library, and then — in five diverging, partially-overlapping
continuations of that same conversation — uses it to blindly score the workflow recommendations produced
by the 32 design-task chats above, for later quantitative comparison of the prompting conditions.

Coverage note: every condition for Claude Sonnet 5 / VAST, plus the `null` chats for Claude/VIS and
GPT/VAST, were read in full. The remaining design-task chats (mostly GPT model runs, and the VIS-domain
runs for Claude beyond `null`) follow the identical shared prompt template confirmed in those readings;
their summaries below are inferred from that template plus each chat's own first/last messages rather
than a full read-through, as noted per row.

## Overview

| File | Topic | Process | Artifacts | Final State |
|---|---|---|---|---|
| Claude_Sonnet_5_VAST_ATWL+library.json | Workflow design by Claude Sonnet 5 for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data), given the ATWL spec, an example, and a 53-workflow library of published workflows pre-formalised in ATWL. | Model confirms it has absorbed all three materials (referencing specific library workflows by name), designs a workflow drawing on and citing relevant library workflows, then lists the source papers it drew approaches from. | Prose + ATWL workflow design citing specific library workflows; a follow-up reference list of source papers. | Complete: design plus citation list delivered; this is also the shared prefix of the longer *_ATWL_2-stage.json branch from the same underlying chat. |
| Claude_Sonnet_5_VAST_ATWL.json | Workflow design by Claude Sonnet 5 for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data), given only the ATWL specification and one worked example (no workflow library). | Model first confirms it has understood the ATWL spec and example, explicitly noting the example's structure shouldn't be copied as a template. It is then given the task and produces a natural-language explanation plus a full formal ATWL specification (artifacts, transforms, loops, actors). | Prose explanation + complete ATWL-formalised workflow specification for the task. | Complete: one workflow designed and fully formalised in ATWL; conversation ends there. |
| Claude_Sonnet_5_VAST_ATWL_2-stage.json | Two-stage enrichment of the ATWL+library workflow design by Claude Sonnet 5 for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data). | Continues the ATWL+library chat (same first 6 turns): after the initial design and citation list, the user uploads the full source papers and asks the model to enrich the (structurally fixed) workflow with method-level detail -- specific methods, parameters, and validation approaches per phase, presented as a table -- while separately flagging any structural changes the papers suggest as non-incorporated 'possible refinements'. | A phase-by-phase methods/parameters/validation enrichment table, plus a separate list of possible structural refinements not incorporated into the design. | Complete: implementation-ready enriched version of the original design delivered; a branch/continuation of the ATWL+library chat. |
| Claude_Sonnet_5_VAST_C6+.json | Extended (4-stage) version of the C6 chat: Claude Sonnet 5 translates its prose design to ATWL, then critiques and revises it. | Continues the C6 chat (same first 8 turns: prose design + ATWL translation). The user then asks the model to reflect on whether working in ATWL revealed gaps in the original design (unspecified loops, missing assess steps, implicit criteria), which the model answers as a numbered list of recommended (not-yet-applied) modifications; finally the user asks it to actually implement the modifications it judges genuine improvements, producing a revised ATWL representation with a changelog of what was/wasn't implemented and why. | A gap-analysis of the ATWL translation, a revised/restructured ATWL workflow implementing selected fixes, and a changelog explaining accepted/rejected/modified recommendations. | Complete: full four-stage arc (prose design -> ATWL translation -> self-critique -> revised ATWL) delivered; the longest branch of the library_prose chat. |
| Claude_Sonnet_5_VAST_C6.json | Second stage of the library_prose chat: Claude Sonnet 5 translates its natural-language workflow for Vastopolis into formal ATWL. | Continues the library_prose chat (same first 6 turns). The user now supplies the ATWL spec/example and asks for a faithful ATWL translation of the earlier prose design -- explicitly not a redesign -- with instructions not to fabricate structure (loops, actor distinctions) beyond what the prose actually described, and to report afterward which structural elements were added/made explicit versus directly translated. | A full ATWL formalisation of the earlier prose workflow, plus a self-report of which structural elements were implicit additions. | Complete: one faithful ATWL translation plus additions note delivered; a branch of the library_prose chat. |
| Claude_Sonnet_5_VAST_C7.json | Alternative second stage of the library_prose chat: Claude Sonnet 5 redesigns the workflow from scratch directly in ATWL rather than translating the prose version. | Continues the library_prose chat (same first 6 turns). The user asks for a fresh ATWL design 'not as a translation' of the earlier prose recommendation, explicitly inviting departures from the original phase structure where ATWL's constructs suggest a better design; the model produces a new ATWL workflow, then on request summarises how it differs structurally from the earlier natural-language version. | A freshly (re)designed ATWL workflow plus a summary of how it structurally departs from the earlier prose design. | Complete: redesign plus departure summary delivered; another branch of the library_prose chat, alternative to C6/C6+. |
| Claude_Sonnet_5_VAST_library_prose.json | Workflow design by Claude Sonnet 5 for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data), given only natural-language summaries of the 53-workflow library (no formal ATWL library, no ATWL spec at this stage). | Model confirms it has absorbed the 53 prose summaries, then designs a workflow entirely in natural language (detailed enough to serve as a design template), citing library workflows by name, and finally lists the source papers drawn on. | Detailed natural-language workflow design (stages, methods, visual encodings) with citations; a source-paper list. | Complete: prose-only design delivered. This chat is also the shared prefix for the C6, C6+, and C7 branches from the same underlying conversation. |
| Claude_Sonnet_5_VAST_null.json | Baseline (no ATWL materials) workflow design by Claude Sonnet 5 for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data). | The model is given only the task description -- no ATWL spec, example, or library -- and asked directly to design an analytical workflow. It replies with a free-form staged plan in ordinary markdown (headings/tables/checkpoints), not in ATWL. | A free-form, non-ATWL staged workflow plan (stages, checkpoints, a hypothesis table). | Complete in a single exchange; ends there (trails off with an unprompted German-language follow-up question offering to elaborate a step with sample code). |
| Claude_Sonnet_5_VIS_ATWL+library.json | Workflow design by Claude Sonnet 5 for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus), given the ATWL spec, an example, and a 53-workflow library of published workflows pre-formalised in ATWL. | Model confirms it has absorbed all three materials (referencing specific library workflows by name), designs a workflow drawing on and citing relevant library workflows, then lists the source papers it drew approaches from. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Prose + ATWL workflow design citing specific library workflows; a follow-up reference list of source papers. | Complete: design plus citation list delivered; this is also the shared prefix of the longer *_ATWL_2-stage.json branch from the same underlying chat. |
| Claude_Sonnet_5_VIS_ATWL.json | Workflow design by Claude Sonnet 5 for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus), given only the ATWL specification and one worked example (no workflow library). | Model first confirms it has understood the ATWL spec and example, explicitly noting the example's structure shouldn't be copied as a template. It is then given the task and produces a natural-language explanation plus a full formal ATWL specification (artifacts, transforms, loops, actors). (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Prose explanation + complete ATWL-formalised workflow specification for the task. | Complete: one workflow designed and fully formalised in ATWL; conversation ends there. |
| Claude_Sonnet_5_VIS_ATWL_2-stage.json | Two-stage enrichment of the ATWL+library workflow design by Claude Sonnet 5 for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus). | Continues the ATWL+library chat (same first 6 turns): after the initial design and citation list, the user uploads the full source papers and asks the model to enrich the (structurally fixed) workflow with method-level detail -- specific methods, parameters, and validation approaches per phase, presented as a table -- while separately flagging any structural changes the papers suggest as non-incorporated 'possible refinements'. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A phase-by-phase methods/parameters/validation enrichment table, plus a separate list of possible structural refinements not incorporated into the design. | Complete: implementation-ready enriched version of the original design delivered; a branch/continuation of the ATWL+library chat. |
| Claude_Sonnet_5_VIS_C6+.json | Extended (4-stage) version of the C6 chat: Claude Sonnet 5 translates its prose design to ATWL, then critiques and revises it. | Continues the C6 chat (same first 8 turns: prose design + ATWL translation). The user then asks the model to reflect on whether working in ATWL revealed gaps in the original design (unspecified loops, missing assess steps, implicit criteria), which the model answers as a numbered list of recommended (not-yet-applied) modifications; finally the user asks it to actually implement the modifications it judges genuine improvements, producing a revised ATWL representation with a changelog of what was/wasn't implemented and why. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A gap-analysis of the ATWL translation, a revised/restructured ATWL workflow implementing selected fixes, and a changelog explaining accepted/rejected/modified recommendations. | Complete: full four-stage arc (prose design -> ATWL translation -> self-critique -> revised ATWL) delivered; the longest branch of the library_prose chat. |
| Claude_Sonnet_5_VIS_C6.json | Second stage of the library_prose chat: Claude Sonnet 5 translates its natural-language workflow for IEEE into formal ATWL. | Continues the library_prose chat (same first 6 turns). The user now supplies the ATWL spec/example and asks for a faithful ATWL translation of the earlier prose design -- explicitly not a redesign -- with instructions not to fabricate structure (loops, actor distinctions) beyond what the prose actually described, and to report afterward which structural elements were added/made explicit versus directly translated. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A full ATWL formalisation of the earlier prose workflow, plus a self-report of which structural elements were implicit additions. | Complete: one faithful ATWL translation plus additions note delivered; a branch of the library_prose chat. |
| Claude_Sonnet_5_VIS_C7.json | Alternative second stage of the library_prose chat: Claude Sonnet 5 redesigns the workflow from scratch directly in ATWL rather than translating the prose version. | Continues the library_prose chat (same first 6 turns). The user asks for a fresh ATWL design 'not as a translation' of the earlier prose recommendation, explicitly inviting departures from the original phase structure where ATWL's constructs suggest a better design; the model produces a new ATWL workflow, then on request summarises how it differs structurally from the earlier natural-language version. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A freshly (re)designed ATWL workflow plus a summary of how it structurally departs from the earlier prose design. | Complete: redesign plus departure summary delivered; another branch of the library_prose chat, alternative to C6/C6+. |
| Claude_Sonnet_5_VIS_library_prose.json | Workflow design by Claude Sonnet 5 for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus), given only natural-language summaries of the 53-workflow library (no formal ATWL library, no ATWL spec at this stage). | Model confirms it has absorbed the 53 prose summaries, then designs a workflow entirely in natural language (detailed enough to serve as a design template), citing library workflows by name, and finally lists the source papers drawn on. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Detailed natural-language workflow design (stages, methods, visual encodings) with citations; a source-paper list. | Complete: prose-only design delivered. This chat is also the shared prefix for the C6, C6+, and C7 branches from the same underlying conversation. |
| Claude_Sonnet_5_VIS_null.json | Baseline (no ATWL materials) workflow design by Claude Sonnet 5 for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus). | The model is given only the task description -- no ATWL spec, example, or library -- and asked directly to design an analytical workflow. It replies with a free-form staged plan in ordinary markdown (headings/tables/checkpoints), not in ATWL. | A free-form, non-ATWL staged workflow plan (stages, checkpoints, a hypothesis table). | Complete in a single exchange; ends there (trails off with an unprompted German-language follow-up question offering to elaborate a step with sample code). |
| GPT_5_6_Terra_VAST_ATWL+library.json | Workflow design by GPT-5.6-Terra for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data), given the ATWL spec, an example, and a 53-workflow library of published workflows pre-formalised in ATWL. | Model confirms it has absorbed all three materials (referencing specific library workflows by name), designs a workflow drawing on and citing relevant library workflows, then lists the source papers it drew approaches from. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Prose + ATWL workflow design citing specific library workflows; a follow-up reference list of source papers. | Complete: design plus citation list delivered; this is also the shared prefix of the longer *_ATWL_2-stage.json branch from the same underlying chat. |
| GPT_5_6_Terra_VAST_ATWL.json | Workflow design by GPT-5.6-Terra for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data), given only the ATWL specification and one worked example (no workflow library). | Model first confirms it has understood the ATWL spec and example, explicitly noting the example's structure shouldn't be copied as a template. It is then given the task and produces a natural-language explanation plus a full formal ATWL specification (artifacts, transforms, loops, actors). (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Prose explanation + complete ATWL-formalised workflow specification for the task. | Complete: one workflow designed and fully formalised in ATWL; conversation ends there. |
| GPT_5_6_Terra_VAST_ATWL_2-stage.json | Two-stage enrichment of the ATWL+library workflow design by GPT-5.6-Terra for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data). | Continues the ATWL+library chat (same first 6 turns): after the initial design and citation list, the user uploads the full source papers and asks the model to enrich the (structurally fixed) workflow with method-level detail -- specific methods, parameters, and validation approaches per phase, presented as a table -- while separately flagging any structural changes the papers suggest as non-incorporated 'possible refinements'. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A phase-by-phase methods/parameters/validation enrichment table, plus a separate list of possible structural refinements not incorporated into the design. | Complete: implementation-ready enriched version of the original design delivered; a branch/continuation of the ATWL+library chat. |
| GPT_5_6_Terra_VAST_C6+.json | Extended (4-stage) version of the C6 chat: GPT-5.6-Terra translates its prose design to ATWL, then critiques and revises it. | Continues the C6 chat (same first 8 turns: prose design + ATWL translation). The user then asks the model to reflect on whether working in ATWL revealed gaps in the original design (unspecified loops, missing assess steps, implicit criteria), which the model answers as a numbered list of recommended (not-yet-applied) modifications; finally the user asks it to actually implement the modifications it judges genuine improvements, producing a revised ATWL representation with a changelog of what was/wasn't implemented and why. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A gap-analysis of the ATWL translation, a revised/restructured ATWL workflow implementing selected fixes, and a changelog explaining accepted/rejected/modified recommendations. | Complete: full four-stage arc (prose design -> ATWL translation -> self-critique -> revised ATWL) delivered; the longest branch of the library_prose chat. |
| GPT_5_6_Terra_VAST_C6.json | Second stage of the library_prose chat: GPT-5.6-Terra translates its natural-language workflow for Vastopolis into formal ATWL. | Continues the library_prose chat (same first 6 turns). The user now supplies the ATWL spec/example and asks for a faithful ATWL translation of the earlier prose design -- explicitly not a redesign -- with instructions not to fabricate structure (loops, actor distinctions) beyond what the prose actually described, and to report afterward which structural elements were added/made explicit versus directly translated. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A full ATWL formalisation of the earlier prose workflow, plus a self-report of which structural elements were implicit additions. | Complete: one faithful ATWL translation plus additions note delivered; a branch of the library_prose chat. |
| GPT_5_6_Terra_VAST_C7.json | Alternative second stage of the library_prose chat: GPT-5.6-Terra redesigns the workflow from scratch directly in ATWL rather than translating the prose version. | Continues the library_prose chat (same first 6 turns). The user asks for a fresh ATWL design 'not as a translation' of the earlier prose recommendation, explicitly inviting departures from the original phase structure where ATWL's constructs suggest a better design; the model produces a new ATWL workflow, then on request summarises how it differs structurally from the earlier natural-language version. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A freshly (re)designed ATWL workflow plus a summary of how it structurally departs from the earlier prose design. | Complete: redesign plus departure summary delivered; another branch of the library_prose chat, alternative to C6/C6+. |
| GPT_5_6_Terra_VAST_library_prose.json | Workflow design by GPT-5.6-Terra for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data), given only natural-language summaries of the 53-workflow library (no formal ATWL library, no ATWL spec at this stage). | Model confirms it has absorbed the 53 prose summaries, then designs a workflow entirely in natural language (detailed enough to serve as a design template), citing library workflows by name, and finally lists the source papers drawn on. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Detailed natural-language workflow design (stages, methods, visual encodings) with citations; a source-paper list. | Complete: prose-only design delivered. This chat is also the shared prefix for the C6, C6+, and C7 branches from the same underlying conversation. |
| GPT_5_6_Terra_VAST_null.json | Baseline (no ATWL materials) workflow design by GPT-5.6-Terra for a public-health epidemic-investigation scenario (find the origin and transmission mechanism of a Vastopolis flu outbreak from geolocated microblogs, a city map, population stats, and weather data). | The model is given only the task description -- no ATWL spec, example, or library -- and asked directly to design an analytical workflow. It replies with a free-form staged plan in ordinary markdown (headings/tables/checkpoints), not in ATWL. | A free-form, non-ATWL staged workflow plan (stages, checkpoints, a hypothesis table). | Complete in a single exchange: a long, thoroughly caveated plan (objectives, data prep, hypothesis table, dashboard mock-up, recommended actions, confirmatory follow-up steps); ends cleanly in English. |
| GPT_5_6_Terra_VIS_ATWL+library.json | Workflow design by GPT-5.6-Terra for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus), given the ATWL spec, an example, and a 53-workflow library of published workflows pre-formalised in ATWL. | Model confirms it has absorbed all three materials (referencing specific library workflows by name), designs a workflow drawing on and citing relevant library workflows, then lists the source papers it drew approaches from. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Prose + ATWL workflow design citing specific library workflows; a follow-up reference list of source papers. | Complete: design plus citation list delivered; this is also the shared prefix of the longer *_ATWL_2-stage.json branch from the same underlying chat. |
| GPT_5_6_Terra_VIS_ATWL.json | Workflow design by GPT-5.6-Terra for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus), given only the ATWL specification and one worked example (no workflow library). | Model first confirms it has understood the ATWL spec and example, explicitly noting the example's structure shouldn't be copied as a template. It is then given the task and produces a natural-language explanation plus a full formal ATWL specification (artifacts, transforms, loops, actors). (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Prose explanation + complete ATWL-formalised workflow specification for the task. | Complete: one workflow designed and fully formalised in ATWL; conversation ends there. |
| GPT_5_6_Terra_VIS_ATWL_2-stage.json | Two-stage enrichment of the ATWL+library workflow design by GPT-5.6-Terra for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus). | Continues the ATWL+library chat (same first 6 turns): after the initial design and citation list, the user uploads the full source papers and asks the model to enrich the (structurally fixed) workflow with method-level detail -- specific methods, parameters, and validation approaches per phase, presented as a table -- while separately flagging any structural changes the papers suggest as non-incorporated 'possible refinements'. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A phase-by-phase methods/parameters/validation enrichment table, plus a separate list of possible structural refinements not incorporated into the design. | Complete: implementation-ready enriched version of the original design delivered; a branch/continuation of the ATWL+library chat. |
| GPT_5_6_Terra_VIS_C6+.json | Extended (4-stage) version of the C6 chat: GPT-5.6-Terra translates its prose design to ATWL, then critiques and revises it. | Continues the C6 chat (same first 8 turns: prose design + ATWL translation). The user then asks the model to reflect on whether working in ATWL revealed gaps in the original design (unspecified loops, missing assess steps, implicit criteria), which the model answers as a numbered list of recommended (not-yet-applied) modifications; finally the user asks it to actually implement the modifications it judges genuine improvements, producing a revised ATWL representation with a changelog of what was/wasn't implemented and why. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A gap-analysis of the ATWL translation, a revised/restructured ATWL workflow implementing selected fixes, and a changelog explaining accepted/rejected/modified recommendations. | Complete: full four-stage arc (prose design -> ATWL translation -> self-critique -> revised ATWL) delivered; the longest branch of the library_prose chat. |
| GPT_5_6_Terra_VIS_C6.json | Second stage of the library_prose chat: GPT-5.6-Terra translates its natural-language workflow for IEEE into formal ATWL. | Continues the library_prose chat (same first 6 turns). The user now supplies the ATWL spec/example and asks for a faithful ATWL translation of the earlier prose design -- explicitly not a redesign -- with instructions not to fabricate structure (loops, actor distinctions) beyond what the prose actually described, and to report afterward which structural elements were added/made explicit versus directly translated. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A full ATWL formalisation of the earlier prose workflow, plus a self-report of which structural elements were implicit additions. | Complete: one faithful ATWL translation plus additions note delivered; a branch of the library_prose chat. |
| GPT_5_6_Terra_VIS_C7.json | Alternative second stage of the library_prose chat: GPT-5.6-Terra redesigns the workflow from scratch directly in ATWL rather than translating the prose version. | Continues the library_prose chat (same first 6 turns). The user asks for a fresh ATWL design 'not as a translation' of the earlier prose recommendation, explicitly inviting departures from the original phase structure where ATWL's constructs suggest a better design; the model produces a new ATWL workflow, then on request summarises how it differs structurally from the earlier natural-language version. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A freshly (re)designed ATWL workflow plus a summary of how it structurally departs from the earlier prose design. | Complete: redesign plus departure summary delivered; another branch of the library_prose chat, alternative to C6/C6+. |
| GPT_5_6_Terra_VIS_library_prose.json | Workflow design by GPT-5.6-Terra for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus), given only natural-language summaries of the 53-workflow library (no formal ATWL library, no ATWL spec at this stage). | Model confirms it has absorbed the 53 prose summaries, then designs a workflow entirely in natural language (detailed enough to serve as a design template), citing library workflows by name, and finally lists the source papers drawn on. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | Detailed natural-language workflow design (stages, methods, visual encodings) with citations; a source-paper list. | Complete: prose-only design delivered. This chat is also the shared prefix for the C6, C6+, and C7 branches from the same underlying conversation. |
| GPT_5_6_Terra_VIS_null.json | Baseline (no ATWL materials) workflow design by GPT-5.6-Terra for a bibliometrics scenario (identify and characterise evolving collaboration communities, their stable cores/leaders, research foci, and merge/split events in the 1990-2024 IEEE VIS paper corpus). | The model is given only the task description -- no ATWL spec, example, or library -- and asked directly to design an analytical workflow. It replies with a free-form staged plan in ordinary markdown (headings/tables/checkpoints), not in ATWL. (This specific chat was not read in full; description follows the same protocol confirmed by reading the the Claude Sonnet 5 / VAST chat under this condition, plus this chat's own first/last messages, which match the shared prompt template.) | A free-form, non-ATWL staged workflow plan (stages, checkpoints, a hypothesis table). | Complete in a single exchange, per the shared null-condition pattern; not read in full. |
| feature_extraction_GPT_5_6_C4.json | Applying the motif/feature vocabulary to score the LLM-generated Vastopolis/VIS workflow recommendations -- a parallel continuation of the same scoring task as feature_extractor_GPT_5_6_Terra.json. | Same setup and protocol as feature_extractor_GPT_5_6_Terra.json (shared 36-turn prefix, same blinded one-at-a-time scoring procedure), but this is a separate fork/regeneration that proceeds through a different point in the batch of recommendations. | A table of per-recommendation binary feature vectors with justifications; ends having just scored a recommendation labelled 'VAST_C4_Terra'. | Partial/in progress: ends mid-batch at the VAST_C4_Terra recommendation; a parallel fork of the same scoring session as the other feature_extract*/feature_extractor* files. |
| feature_extraction_GPT_5_6_Terra_C6.json | Applying the motif/feature vocabulary to score the LLM-generated Vastopolis/VIS workflow recommendations -- another parallel continuation of the same scoring task. | Same setup and protocol as feature_extractor_GPT_5_6_Terra.json (shared 36-turn prefix, same blinded scoring procedure); a separate fork that proceeds through a different point in the batch. | A table of per-recommendation binary feature vectors with justifications; ends having just scored a recommendation labelled 'VAST_C6_Terra'. | Partial/in progress: ends mid-batch at the VAST_C6_Terra recommendation; a parallel fork of the same scoring session as the other feature_extract*/feature_extractor* files. |
| feature_extractor_GPT_5_6_Terra-ATWL.json | Applying the motif/feature vocabulary to score the LLM-generated Vastopolis/VIS workflow recommendations -- the most complete continuation, covering the largest share of the batch. | Same setup and protocol as feature_extractor_GPT_5_6_Terra.json (shared 36-turn prefix, same blinded scoring procedure); this is by far the longest fork (108 turns, roughly 3x the shortest ones), scoring well over 30 individual recommendations before ending on one labelled 'VAST_C7_Terra'. | The most extensive table of per-recommendation binary feature vectors with justifications among the five scoring forks, covering most of the design-task experiment's condition/model/domain combinations. | Most complete of the five scoring forks, though the transcript still ends mid-batch rather than with an explicit wrap-up. |
| feature_extractor_GPT_5_6_Terra-prose.json | Applying the motif/feature vocabulary to score the LLM-generated Vastopolis/VIS workflow recommendations -- a longer parallel continuation covering more of the batch than the -C4/-C6/plain forks. | Same setup and protocol as feature_extractor_GPT_5_6_Terra.json (shared 36-turn prefix, same blinded scoring procedure); this fork runs substantially further (~80 turns, roughly double the shorter forks), getting back around to a second 'VAST_C4_Terra'-labelled recommendation, suggesting it re-covers or extends past where the shorter forks stopped. | A more extensive table of per-recommendation binary feature vectors with justifications than the shorter forks. | Partial/in progress but more complete than the -C4/-C6/plain forks; ends mid-batch. |
| feature_extractor_GPT_5_6_Terra.json | Applying the motif/feature vocabulary to score the LLM-generated Vastopolis/VIS workflow recommendations from the parallel design-task experiment. | Continues the motifs_GPT_5_Terra chat. The user re-uploads the ATWL spec and feature definitions, then feeds in -- one at a time, anonymised and in randomised order -- the workflow recommendations produced by different LLMs/conditions in the separate VAST/VIS design-task experiment (some in ATWL, some prose, some mixed). For each, the model returns a full binary feature vector with one-sentence justifications per feature, applying the same conservative criteria used on the library. Shares a 36-turn prefix (same underlying chat) with feature_extraction_GPT_5_6_C4.json, feature_extraction_GPT_5_6_Terra_C6.json, feature_extractor_GPT_5_6_Terra.json, feature_extractor_GPT_5_6_Terra-ATWL.json, and feature_extractor_GPT_5_6_Terra-prose.json -- all are forks/continuations of one long session. | A growing table of per-recommendation binary feature vectors with justifications; ends partway through scoring, having reached a recommendation labelled 'VAST_null' in the running CSV output. | Partial/in progress: roughly the first several recommendations scored (through the VAST_null condition) before the transcript ends; one of five parallel continuations of the same scoring task. |
| motifs_GPT_5_Terra.json | Defining a binary feature/motif vocabulary to characterise ATWL-formalised visual-analytics workflows. | GPT-5.6-Terra first internalises the ATWL language, then extracts the ATWL source for the 52 published library workflows from an uploaded HTML archive. The user then iteratively works with it to design features going beyond trivial counts: multi-transform 'motifs' with actor-transition patterns (e.g. human->machine->human), then structural indicators, model-role features (created/refined/explored/compared/chosen...), analytical-strategy features (exploratory/confirmatory/comparative/constructive/diagnostic), and iteration-target features (what a loop actually revises). Definitions are repeatedly tightened through discussion (e.g. narrowing what counts as 'model created' to exclude pre-trained/unattended preprocessing), applied to all 52-53 library workflows as two CSV files (binary feature matrix + feature explanations), and finally written up as a user manual. Shares a 36-turn prefix (same underlying chat) with feature_extraction_GPT_5_6_C4.json, feature_extraction_GPT_5_6_Terra_C6.json, feature_extractor_GPT_5_6_Terra.json, feature_extractor_GPT_5_6_Terra-ATWL.json, and feature_extractor_GPT_5_6_Terra-prose.json -- all are forks/continuations of one long session. | workflow_features_binary.csv and workflow_feature_explanations.csv (~45 binary features across 5 blocks, scored for all library workflows) and a Markdown user-manual document describing the encoding scheme. | Complete and thorough: a full, documented coding scheme plus a scored feature matrix for the entire workflow library. This chat is the shared foundation the other 5 files branch from to apply the scheme to new material. |

## Relationships

### Conversation branches

Chats below share the same underlying conversation ID and were forked, regenerated, or continued differently at some point.

- **Shared prefix: 6 turns.** 2 files share this conversation ID and are identical for the first 6 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `Claude_Sonnet_5_VAST_ATWL+library.json` — 6 turns total (ends at the fork point)
  - `Claude_Sonnet_5_VAST_ATWL_2-stage.json` — 8 turns total, 2 unique turn(s) after the fork
- **Shared prefix: 6 turns.** 4 files share this conversation ID and are identical for the first 6 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `Claude_Sonnet_5_VAST_C6+.json` — 12 turns total, 6 unique turn(s) after the fork
  - `Claude_Sonnet_5_VAST_C6.json` — 8 turns total, 2 unique turn(s) after the fork
  - `Claude_Sonnet_5_VAST_C7.json` — 10 turns total, 4 unique turn(s) after the fork
  - `Claude_Sonnet_5_VAST_library_prose.json` — 6 turns total (ends at the fork point)
- **Shared prefix: 6 turns.** 2 files share this conversation ID and are identical for the first 6 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `Claude_Sonnet_5_VIS_ATWL+library.json` — 6 turns total (ends at the fork point)
  - `Claude_Sonnet_5_VIS_ATWL_2-stage.json` — 8 turns total, 2 unique turn(s) after the fork
- **Shared prefix: 8 turns.** 4 files share this conversation ID and are identical for the first 8 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `Claude_Sonnet_5_VIS_C6+.json` — 14 turns total, 6 unique turn(s) after the fork
  - `Claude_Sonnet_5_VIS_C6.json` — 10 turns total, 2 unique turn(s) after the fork
  - `Claude_Sonnet_5_VIS_C7.json` — 10 turns total, 2 unique turn(s) after the fork
  - `Claude_Sonnet_5_VIS_library_prose.json` — 8 turns total (ends at the fork point)
- **Shared prefix: 4 turns.** 2 files share this conversation ID and are identical for the first 4 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `GPT_5_6_Terra_VAST_ATWL+library.json` — 4 turns total (ends at the fork point)
  - `GPT_5_6_Terra_VAST_ATWL_2-stage.json` — 8 turns total, 4 unique turn(s) after the fork
- **Shared prefix: 6 turns.** 4 files share this conversation ID and are identical for the first 6 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `GPT_5_6_Terra_VAST_C6+.json` — 12 turns total, 6 unique turn(s) after the fork
  - `GPT_5_6_Terra_VAST_C6.json` — 10 turns total, 4 unique turn(s) after the fork
  - `GPT_5_6_Terra_VAST_C7.json` — 8 turns total, 2 unique turn(s) after the fork
  - `GPT_5_6_Terra_VAST_library_prose.json` — 6 turns total (ends at the fork point)
- **Shared prefix: 4 turns.** 2 files share this conversation ID and are identical for the first 4 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `GPT_5_6_Terra_VIS_ATWL+library.json` — 4 turns total (ends at the fork point)
  - `GPT_5_6_Terra_VIS_ATWL_2-stage.json` — 8 turns total, 4 unique turn(s) after the fork
- **Shared prefix: 6 turns.** 4 files share this conversation ID and are identical for the first 6 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `GPT_5_6_Terra_VIS_C6+.json` — 14 turns total, 8 unique turn(s) after the fork
  - `GPT_5_6_Terra_VIS_C6.json` — 10 turns total, 4 unique turn(s) after the fork
  - `GPT_5_6_Terra_VIS_C7.json` — 10 turns total, 4 unique turn(s) after the fork
  - `GPT_5_6_Terra_VIS_library_prose.json` — 6 turns total (ends at the fork point)
- **Shared prefix: 36 turns.** 6 files share this conversation ID and are identical for the first 36 turns, then branch -- almost certainly regenerations, forks, or different continuations of one underlying conversation.
  - `feature_extraction_GPT_5_6_C4.json` — 54 turns total, 18 unique turn(s) after the fork
  - `feature_extraction_GPT_5_6_Terra_C6.json` — 54 turns total, 18 unique turn(s) after the fork
  - `feature_extractor_GPT_5_6_Terra-ATWL.json` — 108 turns total, 72 unique turn(s) after the fork
  - `feature_extractor_GPT_5_6_Terra-prose.json` — 80 turns total, 44 unique turn(s) after the fork
  - `feature_extractor_GPT_5_6_Terra.json` — 54 turns total, 18 unique turn(s) after the fork
  - `motifs_GPT_5_Terra.json` — 36 turns total (ends at the fork point)

### Shared prompt / task templates

Chats below open with essentially the same instructions, suggesting the same task template or protocol was reused.

- **8 chats** opening with: _"You have received three files: the ATWL specification, which defines a formal language for representing visual analytics workflows; an example workflow expressed in ATWL; and a document titled ATWL Li..."_
  - `Claude_Sonnet_5_VAST_ATWL+library.json`, `Claude_Sonnet_5_VAST_ATWL_2-stage.json`, `Claude_Sonnet_5_VIS_ATWL+library.json`, `Claude_Sonnet_5_VIS_ATWL_2-stage.json`, `GPT_5_6_Terra_VAST_ATWL+library.json`, `GPT_5_6_Terra_VAST_ATWL_2-stage.json`, `GPT_5_6_Terra_VIS_ATWL+library.json`, `GPT_5_6_Terra_VIS_ATWL_2-stage.json`
- **4 chats** opening with: _"You have received two files: the ATWL specification, which defines a formal language for representing visual analytics workflows, and an example workflow expressed in ATWL. Please read both carefully..."_
  - `Claude_Sonnet_5_VAST_ATWL.json`, `Claude_Sonnet_5_VIS_ATWL.json`, `GPT_5_6_Terra_VAST_ATWL.json`, `GPT_5_6_Terra_VIS_ATWL.json`
- **16 chats** opening with: _"You have received a document titled Workflow Library: Titles and Summaries. This document contains natural-language summaries of 53 visual analytics workflows drawn from the published research literat..."_
  - `Claude_Sonnet_5_VAST_C6+.json`, `Claude_Sonnet_5_VAST_C6.json`, `Claude_Sonnet_5_VAST_C7.json`, `Claude_Sonnet_5_VAST_library_prose.json`, `Claude_Sonnet_5_VIS_C6+.json`, `Claude_Sonnet_5_VIS_C6.json`, `Claude_Sonnet_5_VIS_C7.json`, `Claude_Sonnet_5_VIS_library_prose.json`, `GPT_5_6_Terra_VAST_C6+.json`, `GPT_5_6_Terra_VAST_C6.json`, `GPT_5_6_Terra_VAST_C7.json`, `GPT_5_6_Terra_VAST_library_prose.json`, `GPT_5_6_Terra_VIS_C6+.json`, `GPT_5_6_Terra_VIS_C6.json`, `GPT_5_6_Terra_VIS_C7.json`, `GPT_5_6_Terra_VIS_library_prose.json`
- **2 chats** opening with: _"A public health team in Vastopolis, a metropolitan area of about two million residents, has observed a sharp increase in flu-like illnesses over the past few days, with several deaths believed to be a..."_
  - `Claude_Sonnet_5_VAST_null.json`, `GPT_5_6_Terra_VAST_null.json`
- **2 chats** opening with: _"The IEEE VIS conferences (Vis, InfoVis, VAST, SciVis) have accumulated 3596 papers between 1990 and 2024. Each paper record contains the list of authors, publication year, title, abstract, keywords, a..."_
  - `Claude_Sonnet_5_VIS_null.json`, `GPT_5_6_Terra_VIS_null.json`