Deming PDSA

Use this skill when the user asks for PDSA, PDCA, Deming, continuous improvement, stop-line, andon, standard work, or when work is looping without measurable progress.

Quick Start

1. Create a cycle artifact:
   • nu scripts/new-pdsa-cycle.nu --title "<problem>" --owner "<owner>"
2. Fill in PLAN first; do not execute before gates are defined.
3. Run one controlled DO experiment (single variable change).
4. Run STUDY with proof probes (proveryay verify ... or proveryay batch).
5. Complete ACT: adopt/adapt/abandon and update standard work.

PDSA Contract

Every cycle must include these fields:

• problem_statement: concrete defect or bottleneck
• hypothesis: what change should improve and why
• primary_metric: one numeric metric
• target: numeric target and deadline
• guardrails: max acceptable regressions
• owner: single accountable owner
• timebox: explicit start/end window
• claims: verifiable assertions with shell probes
• decision: ADOPT, ADAPT, or ABANDON
• standard_work_update: exact procedure change

If any field is missing, the cycle is incomplete.

Workflow

PLAN

• Define baseline from observed facts (not assumptions).
• Pick the smallest experiment that can falsify the hypothesis.
• Define stop-line triggers before execution.
• Write claims in machine-verifiable form.

Output:

• plans/pdsa/<cycle-id>.md (plan section complete)

DO

• Execute exactly the planned experiment.
• Keep one-change-at-a-time discipline.
• Record command, timestamp, and output artifact locations.

Output:

• command logs, build logs, artifact paths

STUDY

• Compare baseline vs result for primary metric.
• Verify claims with proveryay.
• Mark each claim PASS, FAIL, or INCONCLUSIVE.

Required rule:

• No verbal completion without probe evidence.

Output:

• study_summary in cycle file with metric delta and proof snippets

ACT

• ADOPT: target met and guardrails hold
• ADAPT: partial progress, adjust one variable
• ABANDON: no gain or guardrail breach

Then update standard work:

• append one concrete change to plans/pdsa/STANDARD-WORK.md

Claim Format

Use this exact shape in cycle files and agent reports:

<claims>
- claim: "L1 build is actively compiling"
  probe: "ssh studio@10.0.2.42 'wc -c /home/studio/bsd-build/logs/buildworld-l1.log'"
  expect: ""
- claim: "Artifact is bootable in QEMU"
  probe: "grep -E 'Booting|login:' /path/to/boot.log"
  expect: "Booting"
</claims>

Stop-Line Rules

Stop execution immediately when any condition is true:

• repeated identical failure signature appears twice in window
• disk guardrail breached
• host unreachable for required lane
• proof probes fail for critical claim

On stop-line:

• open a new PDSA cycle with root-cause hypothesis
• do not retry blindly

Multi-Agent Use

When using subagents:

• assign each agent one lane or one proof surface
• require <claims> block in every agent completion
• treat unverified claims as non-facts
• merge only after at least one independent verifier agent passes

Output Standard

Return results in this order:

1. PLAN summary (hypothesis, metric, target)
2. DO actions (commands + artifacts)
3. STUDY evidence (probe verdicts)
4. ACT decision and standard-work update

Keep it concise and numeric.