Sharpening

“The system gets sharper with use — measurably, weekly.” That is a cadence, not a feature, and this page is the cadence. It is the answer to how the improvement contract actually gets executed, day after day, without anyone remembering to do it.

The cadence

flowchart LR
  W["During work<br/>decisions, facts, ratings written as they happen"] --> N["Nightly<br/>consolidate · digest · recompute ratings"]
  N --> WK["Weekly<br/>review · distil feedback · propose promotions"]
  WK --> Q["Quarterly<br/>rescore the Five"]
  WK -.->|approved| SH["Shared brand lessons<br/>and improved skills"]
  SH -.-> W

During work — write it down now

Agents record decisions, durable facts, and status as they happen, through sanctioned writers scoped to you. Ratings and comments attach to specific artifact versions in the Workbench. An in-session calibration step and an end-of-session wrap-up capture what the conversation taught. This is the only input the rest of the cadence has: consolidation cannot capture what was never written.

Nightly — consolidate and recompute

A scheduled maintenance pass, per agent and per scope:

• Consolidates recent events into a fresh narrative summary — the “what was I doing” an agent reads at session start.
• Digests facts into a ranked top-N for the boot bundle, with the rest still searchable.
• Recomputes skill ratings from the reviews written that day, so the catalogue’s scores are never stale.
• Detects drift — anything that shipped outside its writer path, any capability missing its attribution — and reports it.

Nothing is merged across conversations; a conversation’s summary stays its own.

Weekly — review, distil, propose

Mia’s weekly review reads the week’s ratings, results, feedback, and corrections and produces proposals, not changes:

• Lessons to promote into a brand’s shared knowledge — from a private memory row into the brand’s lessons, with the origin recorded. You approve each one.
• Capabilities to improve — the skills with a rating trend worth acting on, each with the evidence attached.
• Catalogue hygiene — skills never chosen, duplicates, gaps agents flagged as “nothing fits”.

Approval is yours; promotion is deliberate; the previous version and the reasoning are kept.

Quarterly — score the Five

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

Why it is scheduled and not remembered

Every step above is a registered automation with a run record, a scope, and a watchdog (Schedules & triggers). If the nightly pass does not fire, that is a failed run surfaced in the agent’s next boot bundle and on the operator feed — not a silent week of stale summaries. Sharpening that depends on someone remembering to do it is not sharpening.

The failure modes this prevents

• Re-litigating decisions. A choice made in conversation, recorded, and consolidated does not get argued again next month from scratch.
• Drift away from validated approaches. Corrections you make become feedback memories the agent reads before repeating the mistake.
• Improvement by anecdote. A skill changes because its rating trend and results say so — not because one post did well (the contract).
• Private feedback leaking. A note to one agent in one conversation reaches shared brand knowledge only through an approved promotion.

live Writes-as-you-go through sanctioned writers, the nightly consolidation and digest, nightly rating recomputation and drift detection, and the weekly review with owner-approved promotion are live. direction Scored variant comparison for skills, and the quarterly rescore as a standing ritual, are the agreed direction.

Related

The improvement contractThe rules the cadence executes.

What your agent remembersWhere the writes go and what promotion means.

Schedules & triggersThe automations that carry the cadence.

PerformanceThe business half of "sharper".