Skills¶
A Skill explains exactly how to use one Tool correctly. Each callable
<tool.id> has one Skill Article at Skills/<tool.id> titled
Using <tool.id>, with a singular tool: [[Tools/<tool.id>]] binding. Order
across several Tools belongs in the Runbook.
| Field | Meaning |
|---|---|
description |
One-line guidance shown in catalogs. |
obsidience.tool |
The one Tool this Skill binds. |
obsidience.requires |
Exact other Skills this Skill depends on (for example, vault.search requires vault.read, because a search hit must be read before it is relied on). |
Skills are how authority reaches a Tool:
- The Executive's direct Skills (
Agent.skills) supply the native Tool catalog of its conversation. - A Task's Runbook declares the Skills its procedure may load.
A Skill or Tool glowing in the graph means its instructions were supplied to the model, not that it was called; actual calls show as separate operation halos.
Starter Vault Skills¶
37 Skills Articles, generated at build time from obsidience/defaults/vault.
| Skill | Tool | Requires | Guidance |
|---|---|---|---|
| Using application.launch | application.launch |
— | Use the controller's registered identifier. |
| Using camera.observe | camera.observe |
— | Look at physical surroundings or something shown to the camera. |
| Using computer.act | computer.act |
computer.observe |
Choose the intended control from the immediately preceding image. |
| Using computer.hardware | computer.hardware |
— | Answer what this computer can do or how busy it is from its current hardware readings. |
| Using computer.observe | computer.observe |
— | Use the scene to select an exact target; focused targets omit name. |
| Using computer.replay | computer.replay |
— | Answer what just happened on the screens from the rolling replay. |
| Using harness.optimize | harness.optimize |
— | Use the bound optimizer to validate Agent and Runbook instruction improvements against independent cases. |
| Using harness.repair | harness.repair |
harness.status |
Use only the exact recovery candidate from current evidence. |
| Using harness.status | harness.status |
— | Read one current snapshot before diagnosing health. |
| Using lights.set | lights.set |
— | Turn requested registered lights on or off through their verified local power operation. |
| Using lights.visualize | lights.visualize |
— | Enable or stop the owner-requested room music light visualizer with explicit audio delay and honest timing status. |
| Using media.acquire | media.acquire |
— | Register one exact source for an authorized acquisition or media playback request. |
| Using media.library | media.library |
— | Discover music, movies and episodes and select an exact source for the requested destination. |
| Using media.pause | media.pause |
— | Pause requested current browser media using verified native playback state. |
| Using model.benchmark | model.benchmark |
model.inspect |
Benchmark only valid layouts for the exact event model. |
| Using model.configure | model.configure |
model.inspect |
Change only the requested supported fields. |
| Using model.inspect | model.inspect |
— | Inspect the named model's current Source and layouts. |
| Using model.source | model.source |
— | Register only a present, known artifact. |
| Using observations.recall | observations.recall |
— | Recover relevant historical observations without confusing them with current evidence. |
| Using observations.retain | observations.retain |
— | Record only a useful compact observation for this Agent. |
| Using session.unlock | session.unlock |
— | Unlock for the current authorized work through the existing native session lock; trust only its verified result. |
| Using source.handoff | source.handoff |
— | Submit one self-contained finding with exact source:// citations and the necessary dates, qualifications and unresolved limits. |
| Using source.ingest | source.ingest |
— | Preserve direct evidence before synthesizing it. |
| Using source.read | source.read |
— | Read the exact bound Source completely when required. |
| Using task.complete | task.complete |
— | For an ordinary successful answer, status:completed and summary are sufficient. |
| Using task.create | task.create |
— | Select an existing accepted outcome, not a new task name. |
| Using tv.control | tv.control |
— | Use one intent call for TV content, volume, mute, pause and power; navigate with remote keys only as a fallback. |
| Using vault.list | vault.list |
— | Use the narrowest relevant folder and follow the returned row offset. |
| Using vault.maintenance | vault.maintenance |
— | Treat candidates as leads. |
| Using vault.propose | vault.propose |
vault.read |
Read the current accessible target and exact supporting Source first. |
| Using vault.read | vault.read |
— | Use exact references returned by search or the active contract. |
| Using vault.search | vault.search |
vault.read |
Identify the missing fact before lookup. |
| Using vault.validate | vault.validate |
— | Use validation for structural integrity. |
| Using web.fetch | web.fetch |
— | Fetch only the exact relevant URLs. |
| Using web.search | web.search |
web.fetch |
Use a focused query for the identified gap. |
| Using window.activate | window.activate |
— | Use the current exact name and optional Surface to disambiguate. |
| Using window.place | window.place |
— | Use the current destination Surface's integer tile edges, not percentages or screen pixels. |