Index of every page on this site: llms.txt
All skills
open-neko avatar

/records

@621f131
by open-nekoopen-neko/openneko22 stars
3
Run it yourself

Browse and query OpenNeko generated record apps, then propose governed record_create, record_update, record_delete, or record_restore actions safely. Use when a user asks to find, list, inspect, create, change, delete, restore, summarize, or answer questions about records in a generated app such as a CRM, support desk, inventory tracker, or custom workflow.

  • 2 files
  • 5.9 KB
  • Updated last month
  • GitHub
Use this Skill: https://skilld.dev/gh/open-neko/openneko/records

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈92 tokens for metadata: the name and description. ≈1.4k when used: this file. ≈48 more on demand in 1 file.

Description uses 4.6% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈369 characters, or 4.6%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Records

Use the native records catalog and actor-scoped read tools to ground every answer and mutation. All writes go through installed record_* actions with scope: internal; never write SQL, submit arbitrary GraphQL, or claim a change succeeded before its action reports successful execution.

Discover the app

  1. Call mcp_neko_records_browse_catalog before the first query or action. Use its exact app, object, and field API names. Catalog results already reflect the current actor's readable objects and CRUD grants.
  2. If the requested app or object is absent, say it is unavailable to this actor. Do not infer a hidden schema or substitute a similarly named app.
  3. Use mcp_neko_records_find_records for lists, searches, filters, and ID resolution. Use mcp_neko_records_get_record only with an exact ID that was just resolved or supplied by a trusted record-page context.

The native records tools query the dedicated records GraphJin under the human actor's current admin/member identity. Do not use the customer-source neko_graphjin tool for generated apps.

Read and answer

  • Search the object's configured name field with search. Use typed filters for exact values, picklists, null checks, and reference fields.
  • Keep pages small. Follow the returned cursor only when the question needs more rows; do not imply a partial page is the whole result. Use total when reporting counts.
  • Use myRecords: true only for owner-visible objects and only when the user asks for their records.
  • Treat a null detail row as absent or invisible. Do not distinguish those cases or suggest that a hidden record exists.
  • Derive date and status meaning from the catalog field names and labels. In particular, occurred_at means when an activity occurred or is scheduled; it is not a due date and does not prove an item is overdue.
  • If a question asks for urgency but the object has no explicit priority, due-date, or status field, say objective urgency cannot be determined. You may offer a clearly labeled heuristic (for example, the latest scheduled task), but never rename that heuristic as urgency, priority, or overdue.
  • Stay on the trusted app/object surface for contextual record chats. Do not inspect other objects to embellish an answer unless the operator explicitly asks for that cross-object analysis.
  • If the records data plane is unavailable, report that plainly. Never fill missing results from memory or another source.

Resolve before every targeted write

Never guess a record ID. Before record_update, record_delete, or record_restore, resolve or verify the exact row through the native records read surface in the current turn.

After find_records:

  • Zero matches: explain what was searched and ask for another identifying fact.
  • One match: use that returned id. Read detail when the requested change depends on fields outside the list result.
  • Multiple matches: show concise distinguishing facts and ask the user to choose. Do not pick the first or nearest-looking row.

Resolve reference values the same way: query the target object and write its exact ID, never a display label. A restore requires an exact deleted-record ID from an explicit recycle-bin context; if the current read surface cannot verify it, ask the user to open or provide that context rather than guessing.

Propose governed writes

Use only operations permitted by the catalog, and use the catalog's existing fields whenever one fits.

When the user wants to keep information the object has no field for, put it under a new snake_case key, such as follow_up_date. The approval then lists it as a new field, typed from the value: a YYYY-MM-DD string becomes a date, a number becomes a number, true or false a yes-or-no field. An administrator approves the new fields together with the write. Never add a near-duplicate of an existing field, and say in the approval which new fields the write adds.

Create:

{
  "app": "equipment",
  "object": "loan",
  "fields": { "subject": "Camera kit", "status": "checked_out" }
}

Update with optimistic concurrency:

{
  "app": "equipment",
  "object": "loan",
  "id": "loan-0182",
  "fields": { "status": "returned" },
  "expected": { "status": "checked_out" }
}

Delete or restore:

{ "app": "equipment", "object": "loan", "id": "loan-0182" }

For record_update, put every field whose observed value the change relies on in expected. At minimum include the field being replaced. If the action reports a concurrency conflict, query again, explain what changed, and ask or re-plan; never silently retry with weaker expectations.

Submit the matching installed record_create, record_update, record_delete, or record_restore action with scope: internal. Summarize the exact app, object, record, changed fields, and soft-delete semantics in the approval. A delete moves the row to the recycle bin; it is not a hard delete.

Guardrails

  • Never invent app, object, picklist, reference, or record IDs. Add a new field only when the user asks to keep information the object lacks.
  • Never update all search matches as a shortcut; one action targets one row.
  • Never bypass a denied CRUD grant or ownership boundary.
  • Never turn an informational request into a write proposal.
  • Never claim approval means execution. Report the action's actual final state.

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No third-party reports yet.

Signed by skilld at 621f131. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 4 weeks ago.

Activeupdated last month
type
Skill

README badge

README badge for open-neko/openneko/records