Skip to content
Describes how WilsonOS is designed to work.What works today →

Getting sharper

Better next month, and you can see why. The fifth promise is the system gets sharper with use — measurably, weekly. That is easy to claim and easy to fake: an agent that rewrites a shared skill every time one post does well is “improving” only in the sense that it is changing. This page is the set of rules that make sharpening real, and the cadence that carries them out.

Evidence What it is What it is tied to
Your feedback Edits, approvals, ratings and stated preferences The exact version it was given on, and who may see it
Capability tests A skill or workflow compared with a candidate against explicit criteria The version and inputs behind each result
Business results Performance, costs and outcomes from the world A source, a date range, a freshness, and a stated limit on what it proves

They are kept apart because they mean different things. Approved is not the same as effective, and effective once is not the same as better everywhere: a cover you loved may sell no books, and a post may have done well because of timing rather than copy.

  1. Propose. A person or an agent proposes one specific change, names the evidence behind it, and says which of the five promises it lifts.
  2. Test. The candidate runs against the same criteria as the current version. Where a rubric exists the comparison is scored, not eyeballed. For skills this is Skill Showdown: the current skill and the candidate on the same brief, scored, then promoted by a person.
  3. Promote deliberately. If it wins, it becomes current and the previous version, the evidence and the reasoning are kept. If it loses, that is recorded too, so the same idea is not re-proposed next month from scratch.

The word that matters is deliberately. Nothing promotes itself. A strong result is a reason to propose, never a licence to overwrite. Small, obvious upgrades you can approve directly.

None of this works unless every output can be traced back. For each deliverable the system records the task, the run, the skill and its version, the inputs, the version made, and the review it received. That chain turns “this skill feels worse lately” into a comparison of two versions over dated evidence.

Every output you review drives two loops at once.

Life 1 · it teaches the skill
your rating lands on the skill that made it
ratings suggest an upgrade · Skill Showdown proves it before it ships
An artifact
one output you rated
Life 2 · it becomes context
the shared brain grows out of the work
customer research → your ICPscopywriting → your voicepage generation → your destinations
A labelled diagram, not a screenshot.

Life 1: it teaches the skill that made it. Your rating lands on the version you looked at and is credited to the skill that made it, never to the agent. Ratings build a trend; a trend worth acting on becomes a proposed upgrade; Skill Showdown proves it before anything ships.

Life 2: it becomes context for the next one. Some outputs are worth keeping as knowledge. Customer research becomes your ideal-customer profiles; approved copy becomes part of your voice; a published page becomes a destination the next campaign points at. The shared brain grows out of the work instead of being typed in separately.

Four rhythms, each a scheduled routine with a run record and a watchdog.

During work — write it down now. Agents record decisions, facts and status as they happen, scoped to you, and your ratings and comments land on specific versions. This is the only input the rest of the cadence has: consolidation cannot capture what was never written.

Nightly — consolidate and recompute. Recent events become a fresh summary, the “what was I doing” an agent reads when it starts; facts are ranked into a digest for the start-up bundle; skill ratings are recomputed from the day’s reviews; and anything that skipped its proper writer, or lacks its attribution, is reported. Nothing is merged across conversations.

Weekly — review, distil, propose. Mia reads the week’s ratings, results, feedback and corrections and produces proposals, not changes: lessons to promote into a brand’s shared knowledge (each one approved by you, with its origin recorded), skills whose trend is worth acting on (with the evidence), and catalogue hygiene — skills never chosen, duplicates, and gaps where agents said “nothing fits”. The catalogue tightens as well as grows.

Quarterly — score the five promises. Each promise is rated one to five against concrete tests, and the direction is recorded. For promise 5 the questions are literal: are accepted promotions rising, are rejected proposals being re-argued, and is the catalogue smaller and better-routed than last quarter?

Every step is a registered routine with a run record, a scope and a watchdog. If the nightly pass does not fire, that is a failed run that shows up in the agent’s next start-up bundle — not a silent week of stale summaries. Sharpening that depends on someone remembering is not sharpening.

  • Re-arguing decisions. A choice recorded and consolidated is not argued again next month from scratch.
  • Drifting from what works. Your corrections become memories the agent reads before repeating the mistake.
  • Improvement by anecdote. A skill changes because its trend and results say so, not because one post did well.
  • Private feedback leaking. A note to one agent reaches shared brand knowledge only through a promotion you approved.

Sources: input/WilsonOS_Overview.html §07, docs/architecture/skills-and-flows.md (Skill Attribution, Skill Showdown), docs/architecture/memory.md §Sharpening discipline, docs/architecture/agents-runtime.md §Crons, docs/proposals/2026-05-28-north-star.md, WilsonOS Team Brief (prototype), "An artifact has two lives" — absorbed (task 494) · Last verified 2026-10-06