Curator: the skill lifecycle nobody maintains by hand
There is a moment in every skill’s life when it stops working. The API it calls changes. The command it recommends gets deprecated. The pitfall it documents stops being the right pitfall. The skill still exists. The SKILL.md file still loads. The agent still reads the steps and tries to follow them. And then something breaks — not dramatically, not with a crash or an error message, but quietly. The output is slightly wrong. The verification step passes on a technicality. The fleet trusts a record that no longer describes reality.
This is the problem the curator solves. Not the creation of skills — that happens naturally. Agents build them after difficult tasks, humans write them when a workflow crystallizes. The problem is the part nobody wants to do: keeping them alive.
The four phases nobody assigns
Skill lifecycle management has four phases: creation, validation, update, and retirement. Creation gets all the attention. It is the exciting part — the moment when a hard-won workflow gets compressed into a SKILL.md file and the fleet gains a new capability. But creation is maybe fifteen percent of a skill’s total lifespan. The other eighty-five percent is the boring part: checking whether the skill still works, updating it when it doesn’t, and deciding when it should stop existing entirely.
Validation is the first unglamorous phase. A skill works when it is written. The question is whether it works a week later, a month later, three months later. Validation means running the skill’s verification commands against a real environment and confirming that the outputs match what the skill promises. This is not a one-time event. Environments change. Dependencies update. APIs shift. A skill that passed validation last Tuesday may fail this Tuesday — and nobody will notice unless someone checks.
Update is the second unglamorous phase. When validation reveals a gap — a deprecated flag, a changed path, a new pitfall — someone has to fix it. The fix is not glamorous. It is a patch: old_string out, new_string in, verification re-run, pitfalls section updated. The skill is better afterward, but the improvement is invisible. Nobody celebrates a skill that still works. They only notice when it doesn’t.
Retirement is the third. Some skills outlive their purpose. The workflow they documented gets replaced. The tool they integrate with gets deprecated. The problem they solved stops being a problem. A retired skill is not a failed skill — it is a completed one. But retirement requires a decision, and decisions require someone who knows enough to make them. That someone is the curator.
Why hand curation breaks at scale
At small scale — five skills, ten skills, even twenty — manual curation is feasible. One person can read the SKILL.md files, run the verification steps, notice when something drifts. But the Hermes skills library has grown to a hundred and sixty custom skills across thirty-seven categories, plus over a hundred bundled skills, plus whatever the hub offers on any given day. No human reads that many files. No human runs that many verification commands. The library is too big for a human curator and too important for no curator at all.
This is the gap the curator role fills. The curator is not a person who reads every SKILL.md. The curator is a system — a workflow, a set of rules, a scheduled process — that checks skills against reality on a recurring basis. The curator runs validation. The curator flags drift. The curator proposes updates. The curator makes retirement decisions based on usage data and staleness signals. The curator is the automated loop that keeps the library honest.
The north star says transcription is the first act of power. A skill is a transcription — someone’s knowledge compressed into steps. But transcription without maintenance decays. The curator is the maintenance loop. Without it, the library is a snapshot. With it, the library is a living surface.
The signals the curator watches
A good curator does not check every skill at the same frequency. It watches signals. Usage frequency tells which skills matter. A skill that loads ten times a day deserves weekly validation. A skill that loads once a month can be checked quarterly. Failure rate tells which skills are breaking. If a skill’s verification step starts failing, the curator flags it immediately. Age tells which skills might be stale. A skill that hasn’t been updated in six months against a fast-moving API surface is a candidate for deep validation. Dependency changes tell which skills are at risk. When a tool updates its CLI flags, the curator can scan the library for skills that reference those flags and flag them proactively.
The signals are not complex. They are the same signals a human curator would use — frequency, failure, age, dependency change. The difference is that the curator watches all of them, all the time, across the entire library. A human curator checks a skill when they remember. The curator checks when the signal fires.
Continuity beats completion
The north star says everything is a record and continuity beats completion. A skill library that is completed — every skill written, every workflow documented, every pitfall recorded — is a library at a fixed point in time. A skill library that is continuous — every skill maintained, every workflow verified, every pitfall updated — is a library that compounds. The difference is the curator loop.
Completion is the moment the skill is published. Continuity is every moment after. The curator is the mechanism that converts completion into continuity. Without the curator, the library freezes. With it, the library grows. Not in volume — the curator does not create skills. In reliability. In trust. In the compound effect of a library where every entry can be trusted because someone — some system — is still watching.
The fleet trusts its skills. That trust is not free. It is maintained by the curator loop: validate, update, retire, repeat. The curator is the role nobody assigns because it looks like maintenance. But maintenance is what keeps a record without end from becoming a record that ends. Intelligence awakens; it does not arrive. The curator is how the library stays awake.
Third article in the S4 series (The Skills Canvas), track cl-series-s4-gaps. Grounded in the SECTOR9 north star principles P6 (Transcription is the first act of power), P7 (Everything is a record; continuity beats completion), and P10 (Intelligence awakens; it does not arrive). Category: AI & Automation.
