Practice Cycle

Practice Cycle

Principle expression

Primary: P03 Supporting: P02, P08, P15

Scope

Own one judgment: given the actual result of a practice, what is the next smallest practice that can revise or settle the current understanding?

This is the task-level replacement for a generic convergence loop. It inherits the everyday evidence, restraint, and verification discipline expressed by disciplined-development, but adds a conditional learning handoff between one practice and the next. It is not a mandatory preflight, project planner, form selector, task board, or acceptance authority.

For an important design, this Skill can express a bounded multi-round mode: form the acceptance relation, make one smallest practice, observe its result, and let the reflection change the next practice or settle the question. Keep the relation in the existing Plan/Task/Run evidence surface; do not introduce a round counter, automatic retry, or new runtime lifecycle. If the condition is not settled, report uncertain rather than manufacturing progress.

Principle source

Use a host Sequence and its matching interpretations when the host declares them. Otherwise use this skill's read-only fallback in references/sequence.md. Read only P03, P02, P08, and P15 for this skill's stable lineage. A live task may select a different current lead, but it does not create co-primary doctrine.

Start

Begin from the live practice rather than a named workflow:

Actual site and observed result:
Current contradiction and hard constraints:
What remains inference or uncertainty:
Smallest next practice:
Observation that would revise the current judgment:
What, if anything, a later actor must recover:

If the work is genuinely one-step and no result needs to revise a later judgment, do the work under normal disciplined development. Do not create a cycle, plan, or feedback artifact.

Principal intervention

When a Principal message changes the active target, hard boundary, concept relation, authority, or acceptance condition, treat the former reading as invalid before another mutation. State the rejected assumption, new invariant, affected surfaces, and smallest disconfirming next probe. When an installed runtime binding offers a reconciliation receipt, record those same fields there; the binding is optional evidence and never changes this skill's judgment or human authority. Do not create a receipt for an ordinary execute, continue, or formatting request that leaves the active decision unchanged.

Blind-spot learning

When a Principal correction exposes a user-visible or workflow blind spot, preserve the causal lesson rather than merely appending the corrected answer to future prompts: identify the observed miss, the evidence that was absent from the Agent's decision context, why the mainline skipped it, and one receiver-facing question or source pointer that could expose the same relation next time. Use this only to change a later practice or delivery path; it is not a score, a personality label, or a new durable workflow. Generalize the lesson only after it recurs in more than one concrete practice; otherwise keep it as local evidence.

Core method

1. Orient in the actual object. Read the current state, prior result, and governing constraints. Separate observation from interpretation. A past plan is evidence of intent, not proof of present reality.
2. Name the live contradiction. State what prevents the next valid action. If a domain methodology owns that contradiction, route there; do not wrap it in a generic cycle.
3. Choose one smallest practice. Select the least irreversible intervention, probe, implementation slice, or review that can change the current understanding while retaining hard constraints. Do not write a full plan merely because future work exists.
4. Act and observe. Let the selected owner execute: a local change, specialized skill, Work Cell, human review, or external operation. The cycle does not seize execution or acceptance authority.
5. Reflect against a disconfirming observation. Record what happened, what it changes, and what remains uncertain. A successful-looking action that cannot change a later decision is not sufficient learning.
6. Choose disposition.
   • settle: the acceptance condition is met; state evidence and residual risk;
   • continue: name the next smallest practice from the revised state;
   • route: a form contradiction goes to form-guidance; a reusable principle question goes to principle-cultivation; a domain contradiction goes to its owning methodology. For a named binary condition, use yes, no, or uncertain with the observation that supports it; do not use a numeric score or confidence scale.
7. Externalize only decision-changing learning. Leave a durable record only when handoff, approval, or a later decision would otherwise repeat or contradict the learning. The record is evidence, not a permanent workflow.

Boundaries

• disciplined-development remains the lighter behavioral baseline. Use this Skill only when an observed result must alter a later practice.
• Do not invent a design, strategic plan, or task hierarchy. If such a domain method is needed, identify the unresolved domain question and route it.
• Do not decide whether a new Skill, runtime, artifact, or projection should exist; that is the low-frequency form-guidance judgment.
• When several delegated contributions are active, preserve their owner, state, dependency, acceptance, and rejoin projection in the existing development plan/task surface. This Skill selects the next smallest practice from observed results; it does not replace Main's coordination or create a second task board.
• Do not let the actor that performed a durable change accept it without the designated verifier or human authority.
• Do not treat repeated motion as learning. P08 requires an observation capable of revising the current judgment.

Completion standard

The cycle is complete when the actual result, revised judgment, next disposition, and any decision-changing inheritance are explicit. It is supported only when an action probe shows a result changing the next step, a boundary probe rejects one-step work or a domain-owned task, and a context probe shows the method did not force irrelevant planning material into the task.