| name | datacard-generator |
| description | Generate Genesis Mission Datacard v1.2 documentation for scientific datasets by introspecting a directory and filling the structured template. Use when the user asks to create a datacard, dataset card, dataset documentation, dataset metadata, document a dataset, or prepare a dataset for sharing. Supports six capability dimensions (discoverability, accessibility, interoperability, reusability, governed_use, ai_usability) โ pick which ones apply via `supports_*` flags. Also converts MODCON v1 datacards to Genesis v1.2. |
| allowed-tools | Bash(*) Read WebSearch WebFetch |
Generating Datacards
Generate a Genesis Mission Datacard v1.2 by introspecting a dataset
directory and filling both the YAML frontmatter and the markdown narrative
body of the canonical template, prompting the user for fields that
introspection can't infer.
The skill also has a one-shot Convert path for migrating an existing
MODCON v1 datacard to Genesis v1.2.
Validation is driven by the upstream Pydantic model (scripts/genesis_models.py)
applied via scripts/validate_datacard.py. The few warn-level rules the
model cannot express live in references/validation-rules.md.
Workflow
Copy this checklist and check off steps as you go.
Progress:
- [ ] 1. Gather context (dataset_path now; `supports_*` capabilities after step 2, with evidence)
- [ ] 2. Run python3 scripts/introspect.py on the dataset directory
- [ ] 3. Load capability-specific guidance
- [ ] 4. Auto-fill YAML from introspect output
- [ ] 5. Confirm dataset readiness level with the user (optional)
- [ ] 6. Prompt for remaining required fields in batches
- [ ] 7. Cross-check every ORCID/ROR/DOI/OSTI identifier via live APIs
- [ ] 8. Compute filename and write the datacard (YAML + markdown body)
- [ ] 9. Run scripts/validate_datacard.py
- [ ] 10. Address findings; re-validate
- [ ] 11. Present review summary
1. Gather context
Ask for the dataset path now. Do not answer the capability questions
below cold โ first run step 2 (introspect), then come back and propose
each supports_* answer with evidence from what introspection found,
rather than interrogating the user blind before any evidence exists.
- Dataset path โ directory to document. Ask for this now.
- Which capabilities does this dataset support? (ask this after step 2,
once introspection evidence is in hand โ see the note at the end of
step 2). Genesis v1.2 organizes fields into six capability containers.
Propose a default for each, backed by an introspection finding, and ask
the user to confirm Yes/No:
supports_discoverability โ always Yes (schema enforces this). Identification, description, project, release status, contacts, authorship. Minimum core fields.supports_accessibility โ Yes if the dataset is meant to be accessed/shared. Adds access policy, endpoints, dataset scale.supports_interoperability โ Yes if the dataset uses standard formats, structured features, controlled vocabularies, or has documented provenance. Adds data_structure, dates, semantic_layer, provenance, related_resources.supports_reusability โ Yes if the dataset is licensed for downstream use. Adds license, citation, integrity, stewardship, data_quality.supports_governed_use โ Yes if the dataset has access restrictions, PII, export control, or formal review requirements. Adds use_governance, security/sensitivity blocks, compliance, review_provenance_companion.supports_ai_usability โ Yes if the dataset is suitable for AI/ML training, inference, or evaluation. Adds ai_usage policy block.
Each supports_X = Yes triggers a required X block in the YAML and a corresponding set of required fields. Each supports_X = No omits that block entirely.
2. Introspect the directory
Run python3 scripts/introspect.py <dataset_path> and capture the JSON. See
references/introspection-commands.md for
what each output field means.
Now return to the capability questions from step 1. Propose an answer
for each supports_* flag backed by an introspection finding โ e.g., "I
found a LICENSE file and CITATION.cff โ I recommend
supports_reusability=Yes. No train//test/ splits detected โ
supports_ai_usability optional." Confirm each with the user
([Yes/No]) before moving on to step 3.
3. Load capability-specific guidance
Read references/capability-prompts.md (the per-capability prompt sequence).
For deep field-by-field explanation, load the relevant section of
references/genesis_field_guide.md only as needed (the file is large).
4. Auto-fill the data card
Copy the template to the output filename without reading it into
context:
cp <template_absolute_path> <output_path>
Then edit the copied file with targeted sed/patch operations to fill YAML
fields and markdown body regions โ do NOT regenerate the entire file. Read
only the sections you need to modify (typically the YAML frontmatter's ~1,140
lines and specific markdown body sections via sed -n 'START,ENDp'). The
template is over 130KB / ~35K tokens; a naive full read blows the context
budget.
Where to save: inside <dataset_dir>/ by default; ask if the user
prefers elsewhere. Compute the filename via the Filename rule (ยง8 below).
Populate the YAML frontmatter
using this decision table (paths use v2 capability-container structure):
| Genesis field | Auto-fill ifโฆ | Otherwise |
|---|
discoverability.datacard.created_date | always (today, ISO 8601) | โ |
discoverability.datacard.updated_date | if_applicable โ leave blank on initial creation; only set on subsequent edits (see Gotcha below) | โ |
discoverability.datacard.creation_method | always โ "Hybrid" (Title case in v2) | โ |
discoverability.datacard.template_version | always โ "1.2" (matches the vendored schema/template version) | โ |
discoverability.datacard.change_log[0] | always (change_date, datacard_version: "1.2", "Initial creation" or "Converted from MODCON v1") | โ |
discoverability.datacard.filename | computed from discoverability.identification.name (see Filename rule below) | โ |
discoverability.datacard.language | always โ en (override if README is non-English) | โ |
discoverability.datacard.created_by[] | always (AI model first if Hybrid; see references/gotchas.md) | โ |
discoverability.identification.name | from README / CITATION.cff title | prompt |
discoverability.identification.version | from CITATION.cff or default "1.0" | prompt |
discoverability.product_type | never โ ProductTypeEnum (see references/lookup-tables.md) | always prompt |
discoverability.datacard.id | never โ IdentifierClass ({type: local, value: <slug>} for pre-publication is a sensible default) | prompt |
discoverability.dataset_description.dataset_summary | from README first paragraph | prompt |
discoverability.dataset_description.keywords | from README / CITATION.cff | prompt |
interoperability.data_structure.formats | introspect.py formats | prompt (requires supports_interoperability=Yes) |
interoperability.data_structure.features | introspect.py sample_columns (structured form: - name: x objects) | prompt |
interoperability.data_structure.splits | introspect.py splits_detected | leave empty |
accessibility.dataset_scale.record_count / .compressed_bytes | introspect.py | prompt (requires supports_accessibility=Yes) |
reusability.license.spdx_id | introspect.py license_hint | prompt (requires supports_reusability=Yes) |
reusability.license.name | never | prompt whenever the reusability.license block is emitted (Pydantic requires it always, not just when spdx_id=other) |
discoverability.authors[] | from CITATION.cff (use CRediT roles โ see references/lookup-tables.md) | prompt |
reusability.citation.preferred_citation | from CITATION.cff bibtex | prompt at [pub] |
interoperability.provenance.was_generated_by | always prompt (often forgotten) | โ |
discoverability.dataset_description.science_domain | never auto-filled โ closed ScienceDomainEnum (see references/lookup-tables.md) | prompt |
ai_usability.ai_usage.training_use_status / .inference_use_status / .evaluation_use_status | never auto-filled | prompt (Yes | No | Conditional); if Conditional, also prompt for the matching *_use_conditions free-text field |
_repository.* | NEVER โ system-owned | โ |
For each supports_X=Yes, also write supports_X: "Yes" at the top level
of the YAML. The schema (via Pydantic) enforces that the matching X: block
must exist when supports_X=Yes.
ORCID, ROR, DOI, and OSTI award numbers gathered here will be
cross-checked against their public APIs in step 7 โ not optional.
Format validation is handled automatically by the validator.
5. Confirm dataset readiness level (optional)
If the user wants to indicate dataset readiness, ask them to set a level
(1, 2, or 3) as freetext in the datacard narrative. There is no dedicated
dataset_readiness YAML field in Genesis v1.2 โ readiness is expressed
through the combination of supports_* flags that are set to "Yes".
As a heuristic to guide the user (also expressible as a rough count of
supports_* flags set to Yes: 1 typically maps to level 1, 2-3 to level 2,
= 4 to level 3):
- 1 = Discoverable (metadata only;
supports_discoverability=Yes + perhaps accessibility)
- 2 = Interoperable & Reusable (also license, contacts, provenance;
supports_interoperability and supports_reusability)
- 3 = AI-Ready & Trustworthy (also semantic layer, integrity, governed use;
supports_ai_usability or supports_governed_use)
Confirm with the user.
6. Prompt for missing fields
Present auto-discovered values for confirmation. Then ask for unfilled
required fields. Ask 3-5 at a time following the batches in
references/capability-prompts.md. Stop and confirm after each batch.
Key vocabulary changes in v2 (full list in references/lookup-tables.md):
- CRediT taxonomy for
authors[].person.role / .organization.role and
contributors[].person.role / .organization.role (role[] lives inside
the agent sub-block โ see Gotcha below) โ
Conceptualization, Data_Curation, Data_Collection, Formal_Analysis,
Funding_Acquisition, Investigation, Methodology, Project_Administration,
Resources, Software, Supervision, Validation, Visualization,
Writing_Original_Draft, Writing_Review_Editing, Other (16 values). Multi-valued
per author. Replaces our previous creator | contributor | data_collector | curator | publisher | sponsor | other list.
- Title_Case for all enums โ
Published not published, Draft not draft, Hybrid not hybrid, etc.
- Sensitivity is no longer a tier ladder. Use
OverallSensitivityEnum:
Public | Unclassified_Uncontrolled | CUI | UCNI | Classified | Legacy_Controlled | Mixed | Other_Controlled.
- Yes/No/Conditional strings (not Python booleans) for
ai_usability.ai_usage.training_use_status / .inference_use_status /
.evaluation_use_status (renamed from *_use_allowed) and governance
Yes/No fields. If any *_use_status = "Conditional", the matching
*_use_conditions free-text field is required.
science_domain is a closed enum (ScienceDomainEnum, 15 values) โ
see references/lookup-tables.md. No longer free text.
7. Cross-check identifiers via live APIs
Step 7 is not optional. Do not proceed to step 8 (write) without
completing it. Datacards with unverified identifiers can silently
misattribute authorship, funding, or DOIs โ worse than a
slightly-incomplete card.
For EVERY ORCID, ROR, DOI, and OSTI award number in the datacard โ
whether the user provided it or introspection inferred it โ resolve it
against the public API per references/live-enrichment.md. Do not
skip this step. Run it even when the field is already populated.
Enrich every path listed under "Identifier paths to check" in
references/live-enrichment.md. The list is grouped by capability so you
can skip capabilities the user opted out of.
For each lookup:
- Clean match โ silent pass.
- Mismatch โ present both side-by-side, ask the user.
- Datacard incomplete โ API has fields the datacard doesn't; offer to add.
- Does not resolve (404, error) โ warn the user; likely typo.
- Rate-limited โ retry once; if still failing, log and move on.
- Widespread network failure โ if more than 2-3 lookups in a row fail
with connection/network errors (not per-ID 404s), stop enrichment, tell
the user that live enrichment could not run, and mark the datacard as
unverified in the review summary at step 11. Do not proceed to step 8
(write) with silently unverified identifiers.
Present ALL findings from all identifiers to the user as a single
consolidated table (see references/live-enrichment.md ยง Batching
guidance). Do NOT prompt after each individual lookup โ batch the
WebFetch calls, then present one summary.
Use WebFetch. Endpoints in references/live-enrichment.md.
8. Filename + write
Filename rule: genesis_datacard_<snake_case(discoverability.identification.name)>.md,
where snake_case lowercases the name and replaces any non-alphanumeric
run with a single _.
Where to save: inside <dataset_dir>/ by default; ask if the user
prefers elsewhere.
What to write: the canonical template
(references/genesis_v1.0_template.md) with both halves filled:
- YAML frontmatter โ fully populated from the workflow above. Set
supports_discoverability: "Yes" and any other supports_X: "Yes" the
user opted into; omit the corresponding X: block (or set
supports_X: "No") for capabilities the user opted out of.
- Markdown narrative body โ also fully filled. See
references/body-fill-guide.md for the
section-by-section mapping.
Strip all placeholder markup before saving. No [!TODO], <REPLACE:>,
<INSTRUCTIONS:>, <metadata_key:>, ${VARIABLE}, or __VALUE__ tokens
should remain. Verify with:
grep -E '\[!TODO\]|<REPLACE:|<INSTRUCTIONS:|<metadata_key:|\$\{|__VALUE__' <output_file>
9. Validate
Run:
python3 scripts/validate_datacard.py <written_file>
The validator emits structured codes:
MISSING_REQUIRED:<field> โ re-promptBAD_ENUM:<field> โ show enum from references/lookup-tables.md; re-promptBAD_FORMAT:<field> โ show format hint; re-promptINCONSISTENT:<field> โ show conflicting values; ask userSCHEMA_VIOLATION:<field> โ unexpected; investigate
Use --json to get machine-readable output for parsing the findings list.
10. Address findings
Loop steps 6 โ 7 โ 8 โ 9 until --json output has "ok": true. Do not
claim done with un-addressed errors. warn severity findings can stand
in the review summary but errors must be resolved.
On loop-back to step 7, re-enrich ONLY identifiers the user added or
changed in this iteration โ do NOT re-check every identifier from scratch
(see references/live-enrichment.md ยง Re-check-only-changed).
11. Review summary
Present:
- Auto-populated fields (count + brief list)
- User-provided fields (count + brief list)
- Empty /
not_applicable fields (with reason)
- Validator warnings (e.g., filename alignment, workflowโrelease alignment)
- Suggestions for improvement (e.g., add
reusability.data_quality.completeness if
reusability is supported and the field is empty)
Ask if the user wants to revise any section before finishing.
Convert path (MODCON v1 โ Genesis v1.2)
When the user asks to convert an existing MODCON v1 datacard:
- Run
python3 scripts/convert_v1_to_genesis.py <v1_path> --json and capture the JSON report. The report has three lists:
mapped โ fields the converter populated.missing_required โ Genesis fields the converter couldn't map. Iterate over this list and prompt the user.orphans โ v1 fields with no v2 equivalent.
- After prompting the user for the fields in
missing_required, compose
the final YAML inline โ apply the answered values to the converter's
mapped output. Do NOT rerun convert_v1_to_genesis.py after prompts;
the converter only reads the v1 source and cannot re-integrate user
answers.
- Set
discoverability.datacard.creation_method = "Hybrid", template_version = "1.2", and ensure
change_log[0] = {change_date: today, datacard_version: "1.2", summary: "Converted from MODCON v1"}.
- Fill the markdown body using
references/body-fill-guide.md.
- Cross-check every identifier via live APIs (step 7 of the Generate path).
- Compute the filename and write the datacard (step 8 of the Generate path).
- Run the validator (step 9 of the Generate path).
- Present the review summary (step 11 of the Generate path).
Gotchas (read before generating)
For additional gotchas (chronological ordering, ai_model/software
relationship requirement, etc.), see
references/gotchas.md.
-
workflow.state โ release_status. Both are needed (under
discoverability.workflow.state and discoverability.release_status).
Recommended alignments are in references/lookup-tables.md.
-
Sensitivity is no longer a tier ladder. Use
OverallSensitivityEnum (Public | Unclassified_Uncontrolled | CUI | UCNI | Classified | Legacy_Controlled | Mixed | Other_Controlled) on
both discoverability.datacard.sensitivity.overall_sensitivity (the
document) and discoverability.sensitivity.overall_sensitivity (the
dataset). These are independent and often differ โ never default them
to match.
-
primary_id.type should not be doi before a DOI is minted.
Use ark, local, or unregistered for pre-publication states.
-
provenance.was_generated_by is required (when
supports_interoperability=Yes) and often forgotten. Even a one-line
answer adds catalog value.
-
change_log is append-only. On re-runs, add a new entry plus
updated_date bump. Never edit or delete prior entries. The field
name inside each entry is datacard_version (patched locally from an
upstream typo).
-
_repository block is system-owned. Do not populate. Leave it
as-is in the template (the underscore prefix is the parser signal).
-
supports_discoverability is always "Yes". The schema (via
Pydantic) enforces this โ every Genesis datacard has at least the
discoverability block.
-
role[] lives INSIDE the agent sub-block, not on the agent entry
itself. AgentClass (used by created_by, contact,
additional_contacts, authors, contributors, facilities,
related_resources.software|ai_models) has no top-level role slot โ
it is a tagged union of person / organization / ai_model /
software, and each of those four sub-classes carries its own role[]
(CRediT taxonomy). Do not write role: as a sibling of person:.
Correct shape:
- contribution_date: "2026-07-01"
creator:
person:
given_name: "Jane"
family_name: "Doe"
role: [Conceptualization, Data_Curation]
-
science_domain is a closed, quoted-string-with-spaces enum.
discoverability.dataset_description.science_domain and
interoperability.domain_metadata.science_domain both use
ScienceDomainEnum, unlike every other enum in the schema which uses
Title_Case / snake_case tokens. See references/lookup-tables.md
for the full list. Free text is no longer accepted.
-
discoverability.datacard.updated_date is now if_applicable, not
required. Leave it blank on initial creation (a datacard that has
never been updated has nothing to report). Only set it when performing
a genuine update to an existing datacard, alongside a new change_log
entry (see Gotcha #5).
References
