Skills
Your conversation context has valuable skills hidden in them.
Pantheon finds them for you.
Context continuity is not only the archive. A space also accumulates lessons — the preference you had to repeat, the command everyone rediscovers, the failure that hit twice. Those lessons die with each chat unless something writes them down. Pantheon reviews the evidence after a handoff and drafts the ones that earned a place.
Why this is continuity
A skill is the lesson that should change the next agent before it starts.
Memory is fetched when an agent goes looking. A promoted skill’s name and trigger are handed over unasked on resume; its full body is loaded only when the work matches. That is how continuity stays bounded: the next model gets the rule without spending the window on every instruction the project ever learned.
You told the last three agents not to add shadows. Each time the correction landed, worked, and died with the chat. A skill is what stops the fourth agent from making the same mistake.
/pantheon-skills my-projectWaiting on you
no-shadowsAwaiting review3 sessionsUse when adding elevation to any surface in this project.
This project does not use drop shadows.
Flat surfaces are separated with a one-pixel border, never elevation. This came up on the board tab and again on the mockup, and was reverted both times.
Drawn from cursor-a91264d3…#482, cursor-a91264d3…#1435. Every id was checked against the vault before this was saved.
The lesson survives the window
A correction dies with the chat that heard it. A promoted skill rides the next resume, so the next agent — or the next model — starts knowing not to make the same mistake.
Continuity without a longer prompt
Resume advertises each live skill’s name and trigger. The body is fetched only when the work matches. That keeps guidance available without pouring every instruction into context.
Evidence, not a vibe
Every draft cites the turns it came from, and the vault refuses a citation that does not resolve. You can open the moment and decide whether the lesson is real.
The same rule in every agent
Once you promote a skill, Cursor, Claude Code, and Codex all receive the trigger on resume. You write the lesson once. Every later agent can fetch it.
See a correction become a reusable skill.
Pantheon finds the repeated signal and its evidence; the agent drafts the lesson, and you decide whether every future agent should receive it.
How Pantheon generates them
When a handoff adds turns, the agent reviews recurring signals and batch-saves only justified, non-repeating lessons as drafts. Run /pantheon-skills my-project to review again at any time. Approval still happens in the UI, on the space page.
What it reads: the corrections you had to repeat, the commands that recur, and any failure that hit more than once. It also checks AGENTS.md and README, and will not write something they already say.
Corrections you had to repeat
A pushback in one conversation is a preference. The same pushback in a later one is a rule nobody wrote down — and that is the clearest signal an archive still holds, because what you typed is always kept.
Procedures that keep coming back
A command run across several sessions is either already documented or is something everyone rediscovers. The review checks your AGENTS.md and README and says which — a lesson already written down does not need a second home.
Failures that recurred
The same command failing the same way in separate conversations, then being worked through each time. Detected without a model where the agent records what its tools returned, which is not most of them.
And what is in the room
The conversation being reviewed, which is the only place the errors an agent hit are still visible. This is usually the strongest evidence and the reason the command exists at all.
The awkward part shaped the design. A skill should be written from evidence — the command that failed, the output that said why — and most agents never write that down. Cursor records which tools it called and with what, but not a single one of the results: 26,000 archived turns here yielded 4,732 tool calls and no outputs at all.
So the only witness to what actually went wrong is the agent that was there when it did. That is why a review is a judgement made inside a live conversation rather than a scan run over the archive afterwards. The archive contributes what it reliably keeps — what you typed, and what got run — and the agent contributes what it saw.
Against a model’s eagerness to be useful there are three checks. It is told, in the tool itself, that writing nothing is the ordinary outcome. Every claim has to cite the turns it came from, and the vault refuses a citation that does not resolve to a real turn. And nothing it writes is a skill yet: it is a draft, inert, until you read it and promote it.
What keeps the list worth reading
- You approve every one
- The review writes drafts without interrupting you. None reaches an agent until you promote it on the space page; reject the rest there.
- Every claim is sourced
- A skill names the turns it came from, and the vault refuses to save a citation that does not resolve. You can open the moment and check it.
- Twelve, at most
- Each live skill’s name and trigger ride every resume, so even lazy loading has a context cost. The cap is the feature — a list nobody reads is worse than no list.
- Then every agent has it
- Promoted skill triggers arrive with the resume in whichever model you open next. The body is fetched only when the work matches.
Start with one conversation
Choose a vault, hand off a conversation, and let Pantheon draft the lessons that keep repeating.