---
name: improve-workflow
description: Convert observed failures and successful methods into tested, retrievable skills, tools or routing changes, then verify reuse and retire ineffective guidance.
---

# Improve the factory, not just the journal

Use after a meaningful reusable signal, or for a selected factory-improvement task.
A brief consideration is enough when there is no useful observation. Do not append
four empty answers after every task.

## Capture and place

Record the observed episode, conditions and evidence. Include successful shortcuts,
not only failures. Find an existing matching lesson before duplicating it. Classify
the improvement at the right owner: project fact, user preference, reusable skill,
mechanical script, model-routing hint, or genuinely changed role responsibility.
A local mistake does not automatically justify a global rule.

## Change and compare

Make one small testable hypothesis: what changes, why it should help, and what
unchanged acceptance will establish success. Within authorized local improvement
scope, have Worker implement it and an independent reviewer/check test it.
Use the original failure and a few contrasting/regression tasks where possible.
For a consequential design, use council; for a tool's actual usability, use Testers.

Compare accepted outcomes, relevant cost, repairs and failure modes. Do not claim
improvement merely because a prompt is shorter or mandatory verification vanished.
Do not optimize a metric by weakening the task or granting the factory new authority.
Candidate tests and model-produced judgments are evidence of their own type, not
proof of successful real-world execution.

## Retrieve and reuse

Publish the method with a clear description of when it applies, a small procedure,
evidence links and limitations. Add it to the skill catalog or the existing native
learning system so a later relevant task can find it. Record subsequent applicable
reuse separately from the original test. No relevant later task means pending
reuse, not failed learning and not invented success.

The learning index may use: observed -> experimenting -> verified -> reused,
with rejected/retired outcomes. These labels express evidence states rather than
a required bureaucracy. Retain the original observation; link actual changes.
Use existing task records for detail instead of a second database.

## Simplify and retire

A successful improvement can delete a rule, merge duplicate skills, replace prose
with a script, remove a pointless agent call or change a model allocation. Retire a
method whose assumptions no longer hold. Preserve earlier evidence and explicit
user constraints; reversible local trials must not erase unrelated shared work.
Hermes-native learning may own storage and triggering when attested, but still
needs retrievability, validation and relevant reuse. Do not duplicate its loop.