> workflow / continuous-improvement

Continuous Improvement

Convert incidents, retros, and metrics into updated prompts, skills, evals, and debt paydown — then feed the next idea cycle.

What should I do next?

Complete the steps and evidence below, clear the human gates, then loop improvements back into prompts, skills, and evals.

When to use

  • A release or incident produced learnings that must not stay in chat
  • Eval golden sets or prompts drifted from production reality
  • Technical debt from AI velocity needs scheduled paydown
  • You are preparing the next initiative with better harness defaults

Step-by-step

  1. Collect inputs: incident actions, retro notes, cost/quality metrics, review themes.
  2. Run knowledge-capture.prompt; land updates in skills/prompts/docs — not wikis alone.
  3. Prioritize technical-debt.prompt items with expiry; schedule paydown slices.
  4. Improve evals that missed real failures; version prompts with regression suites.
  5. For legacy drag, apply legacy-modernization or migration patterns deliberately.
  6. Propose standard/harness changes via PR; pilot before org-wide rollout.
  7. Close the loop by feeding the next Idea brief with constraints learned.

Decisions to make

  • What becomes a standard vs a team-local convention
  • Debt paydown vs new feature capacity split
  • Prompt/skill version bump policy
  • Whether a migration initiative is warranted

Evidence to collect

  • Merged updates to skills/prompts/evals
  • Debt backlog with owners and dates
  • Eval regression results on prompt changes
  • Retro actions closed or explicitly deferred

Role-specific next actions

  • PM: Reserve capacity for debt and harness improvements — not only features.
  • Architect: Promote cross-cutting learnings into ADRs and platform defaults.
  • Engineer: Land skill/prompt/eval PRs; refuse tribal knowledge.
  • Engineering Manager: Run the pilot → adopt cycle for new patterns.

Inputs

  • Incidents and PIR actions
  • Retro notes and review themes
  • Production quality/cost metrics

Outputs

  • Updated skills, prompts, and docs
  • Debt and eval improvement backlog
  • Candidate standard changes for pilot

Required prompts

Required skills

  • Documentation

    Keeps learnings in-repo and reviewable.

  • Evaluation

    Evolves golden sets and gates from production misses.

Required MCP servers

  • GitHub MCP

    Land improvement PRs and track harness changes.

  • Jira MCP

    Optional: track debt paydown work.

Tooling & handbook

Recommended agents & patterns

Human approval gates

  • Standard/harness changes reviewed before org-wide rollout
  • Eval/prompt changes require regression green before merge

Automation opportunities

  • Prompt version bumps tracked in git
  • Eval regression on prompt/skill changes
  • knowledge-capture.hook.sh after SEV close

Enterprise best practices

  • Pilot → adopt cycle for new patterns
  • Contribute proven improvements back to AIES / internal standards
  • Measure harness ROI (defect escape, token cost, lead time)

Common mistakes

  • Tribal knowledge only in chat
  • Never updating golden evals after production misses
  • Declaring "process improvements" without merged artifacts