TL;DR
Obelisk already helps agents revisit their history and gives people a place to browse it. A near-term product direction is to help turn repeated, proven ways of working in that history into reusable agent Skills — and, later, to help discover Skill candidates that have emerged before anyone thought to name them.
The idea
The value of history is not limited to remembering or summarizing what happened.
When the same kind of work happens repeatedly, people and agents try approaches, make corrections, preserve constraints, and gradually settle on methods that continue to work. Once a pattern has survived more than one task and proved useful in practice, it may deserve to become a Skill.
The strongest Skills should not have to begin as generic instructions written from scratch. They can emerge from evidence: what was repeatedly attempted, what failed, what changed after feedback, and which parts remained valuable the next time around.
History is evidence, not a command. Skill generation should help turn that evidence into something inspectable and reusable.
Recaps and Skills
Recaps and Skills are related, but they distill different things:
- A Recap distills a stretch of work: what happened, what changed, and what was learned.
- A Skill distills what still holds across repeated work: what should be possible to do again, how to approach it, and where its boundaries are.
A Recap helps us understand the past. A Skill lets part of that past become a capability for the next task.
Current product boundary
Obelisk currently makes multi-provider session history queryable by agents and browsable by people. It also has a user-triggered Recap flow and can present existing Skill invocations inside a session timeline.
It does not yet provide a product flow for synthesizing a new Skill from historical work or for detecting recurring cross-session patterns as Skill candidates.
Direction preview
Near term: assisted Skill generation
Obelisk should help a person and their agent review relevant history and shape recurring practice into a candidate Skill. A useful result should make it possible to inspect:
- the recurring task or intent the Skill addresses;
- the method that remained useful across attempts;
- important constraints, corrections, and failed approaches;
- the historical evidence that supports the draft;
- what still needs human review before the Skill is saved or used.
The goal is not to silently turn every summary into an instruction. It is to make grounded Skill authoring easier when repeated experience has already done much of the discovery work.
Later: Skill discovery
Further ahead, Obelisk could help surface Skill candidates proactively: clusters of related work, repeated corrections, or stable procedures that have appeared across sessions but have not yet been named or packaged.
These should be explainable suggestions, not silently activated behavior. The exact interaction, confidence model, supported Skill formats, and review boundary remain open questions.
Why open this now
This issue is a community preview and a place to shape the direction before it becomes a detailed implementation proposal. It is not a commitment to a specific UI, architecture, automation level, or release date.
We would especially like to hear:
- What repeated work would you most want Obelisk to help turn into a Skill?
- What evidence would convince you that a pattern is mature enough to preserve?
- Which corrections, failures, and provenance should travel with a generated Skill?
- How much should Obelisk draft automatically, and what should always require explicit review?
- Which Skill formats or agent ecosystems should matter first?
- When would proactive Skill discovery feel helpful, and when would it become noise?
Direction of travel
Let the past become not only something you can remember, but something your agent can call on the next time it matters.
This is one small part of the product improvement Obelisk is meant to make possible.
TL;DR
Obelisk already helps agents revisit their history and gives people a place to browse it. A near-term product direction is to help turn repeated, proven ways of working in that history into reusable agent Skills — and, later, to help discover Skill candidates that have emerged before anyone thought to name them.
The idea
The value of history is not limited to remembering or summarizing what happened.
When the same kind of work happens repeatedly, people and agents try approaches, make corrections, preserve constraints, and gradually settle on methods that continue to work. Once a pattern has survived more than one task and proved useful in practice, it may deserve to become a Skill.
The strongest Skills should not have to begin as generic instructions written from scratch. They can emerge from evidence: what was repeatedly attempted, what failed, what changed after feedback, and which parts remained valuable the next time around.
History is evidence, not a command. Skill generation should help turn that evidence into something inspectable and reusable.
Recaps and Skills
Recaps and Skills are related, but they distill different things:
A Recap helps us understand the past. A Skill lets part of that past become a capability for the next task.
Current product boundary
Obelisk currently makes multi-provider session history queryable by agents and browsable by people. It also has a user-triggered Recap flow and can present existing Skill invocations inside a session timeline.
It does not yet provide a product flow for synthesizing a new Skill from historical work or for detecting recurring cross-session patterns as Skill candidates.
Direction preview
Near term: assisted Skill generation
Obelisk should help a person and their agent review relevant history and shape recurring practice into a candidate Skill. A useful result should make it possible to inspect:
The goal is not to silently turn every summary into an instruction. It is to make grounded Skill authoring easier when repeated experience has already done much of the discovery work.
Later: Skill discovery
Further ahead, Obelisk could help surface Skill candidates proactively: clusters of related work, repeated corrections, or stable procedures that have appeared across sessions but have not yet been named or packaged.
These should be explainable suggestions, not silently activated behavior. The exact interaction, confidence model, supported Skill formats, and review boundary remain open questions.
Why open this now
This issue is a community preview and a place to shape the direction before it becomes a detailed implementation proposal. It is not a commitment to a specific UI, architecture, automation level, or release date.
We would especially like to hear:
Direction of travel
Let the past become not only something you can remember, but something your agent can call on the next time it matters.
This is one small part of the product improvement Obelisk is meant to make possible.