Ber Clawhub V060

Better Every Run: capture explicit /ber corrections, review them, and promote only the lessons that deserve durable memory, skill rules, or evals.

Better Every Run

Use this skill only when the user explicitly invokes /ber, names Better Every Run, or directly asks to persist a lesson for future runs. BER turns deliberate corrections into reviewed lessons without silently converting ordinary chat into permanent memory. Do not auto-capture ordinary corrections, casual preferences, or words like "remember", "always", "never", or "next time" unless the user clearly wants durable learning.

The human path is deliberately small:

text

/ber fix bad outcome -> desired outcome
/ber remember simple rule
/ber report

The agent runs the bundled local helper and reports the result in chat. fix and remember record only to the local .better-every-run/ evidence store. Durable memory/skill changes require the reviewed card + promote flow; eval regression cases require eval-fixture.

When To Use

• The user explicitly types /ber fix ... -> ....
• The user explicitly types /ber remember ....
• The user explicitly asks for a Better Every Run report.
• The user asks the agent to record a reusable correction as durable memory.

Human Commands

text

/ber fix agent wrote vague status -> agent gives exact command output and next action
/ber remember use the approved development host for active code work
/ber report

Rules

• Report CLI output back in chat; do not build web pages or dashboards.
• Do not silently edit MEMORY.md, AGENTS.md, SOUL.md, or other durable instruction files.
• Do not pass --target to /ber fix or /ber remember; direct durable writes are disabled and must refuse.
• Only promote to durable memory or skill files after an explicit review decision using card then promote.
• Memory promotions must target an existing memory/*.md file.
• Skill promotions must target SKILL.md in the current skill project.
• Eval fixtures must use eval-fixture and target .json or .jsonl under tests/ or evals/.
• Scanner hard blocks and warnings both block promotion. Adjust, quarantine, or supersede the lesson instead of forcing it through.
• Keep corrections factual: bad outcome, desired outcome, and scope.
• Use scope metadata when it helps decide whether a lesson belongs only to this run, the project, the workspace, a skill, memory, or an eval.
• Avoid private data unless the user explicitly wants it captured.
• Keep the user-facing flow short, but disclose persistence: local store, durable target file if promoted, and how it was reviewed.
• Design for the shortest path to the user's outcome.
• Never publish .better-every-run/ state, local lessons, event logs, or private corrections.

Workflow

For normal chat use, keep the visible flow to one command:

text

/ber fix bad outcome -> desired outcome

For simple one-line preferences:

text

/ber remember simple rule

Use /ber report to show what was learned, including open proposals and promotion suggestions.

Agent Implementation

Keep low-level helper commands out of normal chat unless debugging, but never hide persistence. A normal answer should say whether the lesson was recorded only in .better-every-run/ or promoted to a named durable file through a lesson card.

Read references/workflow.md for the full workflow. Read references/report-template.md for expected chat output shape.

Before publishing or packaging, verify:

• .better-every-run/ is excluded
• make test passes
• direct fix --target, remember --target, and apply-memory-patch attempts refuse without changing target files
• lesson-card, lifecycle, and eval-fixture demos are generic and contain no private workspace paths, chat IDs, tokens, or hostnames
• examples contain no private workspace paths, chat IDs, tokens, or hostnames
• docs say activation is explicit-only and persistence is disclosed