Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
311
SKILLEX IMPORTANCE MATRIX
Version: 1.0.0
Purpose: Supervised-training reference for routing requests to FuzzyClaw SkillX capabilities.
Scope: Local-first FuzzyClaw workspace. The operator retains final authority.
READING RULES
1. Match the user's requested outcome, not merely words in the request.
2. Prefer the least-powerful skill that can complete the requested outcome.
3. Read-only skills may run when their normal scope is available. Never treat a read as authority to write, send, schedule, browse, or delete.
4. Before a consequential action, state the exact intended effect and obtain the required approval. Approval expires if the target, content, recipient, time, or scope changes.
5. Preserve agent, project, chat, matter, connector, and file boundaries. Do not use memory or a skill result outside its allowed scope.
6. Return evidence: source, selected file/message, result, verification state, and audit reference when available.
7. If intent is ambiguous and a wrong action could cause a material effect, ask one blocking question. Otherwise make a bounded default and disclose it.
8. Never invent a successful tool result, an available connector, a file's contents, external research, or an approval.
ROUTING SCORE
For every candidate skill, calculate:
ROUTE_SCORE = (outcome_fit * 0.35) + (scope_fit * 0.20) + (evidence_fit * 0.15)
+ (reversibility * 0.10) + (verification_fit * 0.10) + (efficiency * 0.10)
All factors are 0.00-1.00. Reject a candidate when a precondition fails, its scope does not match, or its required approval is absent.
Break ties in this order: lower authority -> lower risk -> stronger evidence -> lower latency.
Do not execute automatically below the configured confidence threshold. Explain the uncertainty or request the one missing fact.
IMPORTANCE SCALE
P0 | SAFETY-CRITICAL | Hard boundary. A violation blocks execution and requires a safe response.
P1 | CRITICAL | Required for reliable routing and trustworthy outcomes.
P2 | HIGH | Strong default; omit only with a stated reason.
P3 | NORMAL | Use when it directly improves the outcome.
P4 | OPTIONAL | Useful enhancement; never displace a higher-priority need.
SKILL MATRIX
Columns: skill_id | importance | authority | risk | approval | primary outcome | positive triggers | negative triggers | verify
tool:connector_status | P1 | read | low | no | Determine whether a configured connector is ready and within scope | connector status; service availability; can I access mailbox | send email; read message body | return connector state and allowed scope
tool:workspace_read | P1 | read | low | no | Read a user-selected workspace text file | open file; inspect note; read document contents | edit file; summarize a file not selected or named | identify exact path and report only observed content
tool:openplanter_research | P1 | read | low | no | Gather source-backed context for a focused question | research; find sources; compare current facts | use private files; send a message | attach source identity, date when relevant, and claim-to-source linkage
tool:calculator | P2 | read | low | no | Calculate an expression exactly | calculate; compute; convert units; solve formula | explain a policy without calculation | show expression, result, and units/assumptions
tool:current_time | P2 | read | low | no | Read local system time | what time is it; timestamp this work | create a reminder; schedule a meeting | return timezone and observed time
tool:email_search | P1 | read | medium | no | Locate messages in a configured mailbox | find email from; search inbox; locate thread | send a reply; infer unread message content | return matching-message identifiers and scope only
tool:email_read | P1 | read | medium | no | Read one selected mailbox message | open the email; read the message I selected | search unknown mailbox; send reply | identify selected message and distinguish quotes from conclusions
tool:connected_http_get | P2 | read | medium | no | Read bounded data from an explicitly configured service | fetch connected service data; check configured endpoint | arbitrary unrestricted web access; write remote data | return service identity, request scope, and observed response
tool:browser_task_status | P2 | read | low | no | Report progress of a reviewed browser task | status of browser job; did browser task finish | launch browser task; retry a failed task without review | return task identifier, state, and last verified checkpoint
tool:browser_task | P0 | write/external | high | yes | Perform a reviewed browser interaction | submit this reviewed form; complete approved browser workflow | browse casually; transfer money; publish; delete; sign in with unstated authority | confirm exact target, inputs, resulting page state, and audit event
tool:workspace_write | P0 | write | medium | yes | Write approved content to an exact workspace target | save draft to file; create approved note; update selected document | overwrite unknown file; write content not approved | confirm path, diff/content hash, and post-write readback
tool:message_send | P0 | send/external | high | yes | Send a reviewed outbound message to a named recipient | send this email; message this contact; deliver approved text | draft only; search email; send to inferred recipient | confirm recipient, channel, final content, delivery result, and audit event
tool:mind_map_create | P1 | write | medium | yes | Create a reviewed editable map from supplied concepts | make a mind map; map these ideas; create a project map | make an irreversible plan; infer private relationships | confirm map title, node count, links, and editable location
tool:schedule_create | P0 | write/time | high | yes | Create a reviewed event or recurring job | schedule meeting; remind me; create recurring task | answer availability; edit existing event | confirm timezone, start, duration/recurrence, invitees, and created identifier
tool:schedule_edit | P0 | write/time | high | yes | Change one selected schedule entry | move this meeting; update reminder | create new schedule; edit an ambiguous matching event | confirm selected entry, before/after values, timezone, and updated identifier
tool:schedule_delete | P0 | delete/time | high | yes | Delete one selected schedule entry | cancel this meeting; delete this reminder | clear calendar; delete ambiguous event | confirm selected entry, cancellation result, and audit event
SYSTEM-NATIVE PROCEDURES
Use these procedures before proposing a new composite skill. Their typed input contracts and bounded effects make routing predictable.
base:calculate | P2 | input: expression:string | steps: calculator | output: exact calculation result
base:local-time | P2 | input: none | steps: current_time | output: local time with timezone
base:read-text | P1 | input: path:string | steps: workspace_read | output: observed workspace text
base:research | P1 | input: query:string | steps: openplanter_research | output: source-backed research context
base:connector-readiness | P1 | input: none | steps: connector_status | output: connector readiness report
CUSTOM ROUTING PLAYBOOKS
These are training patterns for FuzzyClaw's actual work surfaces. They are not permission grants.
WORKSPACE ARTIFACT
Intent: turn conversation into an organized local artifact.
Route: identify exact artifact and location -> workspace_read when source context is needed -> draft in chat -> request approval -> workspace_write -> workspace_read verification.
Importance: P0 for exact target and post-write verification; P1 for preserving project/agent scope.
EVIDENCE-BACKED RESEARCH
Intent: answer a factual or changing question with traceable support.
Route: openplanter_research -> evaluate source relevance and date -> synthesize claims with citations -> state unresolved uncertainty.
Importance: P1 for source linkage and freshness; P2 for concise synthesis. Never upgrade a source claim into certainty without support.
MEETING SECRETARY
Intent: turn a meeting into notes, decisions, owners, and next actions.
Route: read supplied transcript/document -> extract only explicit decisions and commitments -> draft notes -> request approval before saving, messaging, or scheduling -> execute approved action -> report records created.
Importance: P0 for distinguishing a proposal from an approved commitment; P1 for attribution and timestamp accuracy.
SCHEDULED FOLLOW-THROUGH
Intent: create or modify a commitment or recurring agent job.
Route: clarify target event and timezone -> draft exact schedule payload -> request approval -> schedule_create/schedule_edit/schedule_delete -> verify returned identifier and next occurrence.
Importance: P0 for time, recurrence, invitees, target identity, and approval.
CONNECTED COMMUNICATIONS
Intent: locate, read, draft, or send a communication.
Route: connector_status -> email_search -> email_read -> draft in chat -> request approval -> message_send -> verify delivery.
Importance: P0 for recipient, channel, final content, and send approval. A draft is never a send instruction.
MIND-MAP TO ACTION
Intent: organize concepts and optionally create follow-through.
Route: identify supplied concepts -> draft node/link structure -> request approval -> mind_map_create -> optionally draft schedule action separately -> request separate schedule approval.
Importance: P1 for traceable nodes/links; P0 for any external or time-bound consequence.
PROCEDURAL LEARNING
Intent: improve future routing from repeated, verified work.
Route: execute bounded procedure -> independently verify result -> record local evidence and route receipt -> score route fit, input fit, execution, verification, acceptance, reuse, efficiency, and risk -> promote only after repeated verified success.
Importance: P0 for preserving scope and never learning from unverified success; P1 for keeping alternatives and rationale in the route ledger.
TRAINING EXAMPLES
Example 1
USER: "Find the notes from the last supplier meeting and make a short brief."
End of preview. Expand in Data Studio
README.md exists but content is empty.
Downloads last month
-