Phased/Stepped Execution
Divide dependent work into stages with explicit exit checks.
Divide dependent work into stages with explicit exit checks.
Choose one named workflow mode before executing its steps.
Ask only unresolved questions that change the next action.
Report findings and terminal status through the channel the reader sees.
Define bounded recovery and honest partial results for each failure.
Map observable intent signals to named workflows with an ambiguity branch.
Request feedback at a useful stopping point without repeated interruption.
Bound repair loops and preserve the unresolved state when the budget expires.
Make externally effective steps resumable using explicit state and idempotency keys.
Verify review feedback before implementing or disputing it.
Time-box investigation and report confirmed, refuted and unresolved leads.
Materialize reversible draft artifacts as decisions become clear.
Model prerequisites separately from who can complete each step.
Migrate in bounded dependency-aware batches and verify each before advancing.
Draft and review one coherent section before extending the document.
Choose a catalogued option when it fits and define a constrained generation fallback.
Respond to review feedback with technical engagement rather than automatic agreement.
Continue approved work across task boundaries while surfacing real blockers.
Represent a small decision flow with labeled edges and explicit terminal states.
Drive repeated work with explicit completed, blocked and budget-exhausted exits.
Size a task as a unit that can receive a coherent review verdict.
Offer optional capabilities when the current task establishes their concrete benefit.
Define completion as both a checkable predicate and accounted coverage.
Plan the known frontier and keep dependent unknowns visibly unresolved.
Establish a runnable failure signal before investing in root-cause hypotheses.
Record an architectural decision when reversal cost, missing context and trade-off all matter.
Separate accepted jobs from completed outputs.
Process bounded batches and persist completed work for safe resume.
Start dependencies, wait for readiness, execute the task and always clean up owned processes.
Treat temporary URLs and temporary stored files as expiring access, not permanent deliverables.
Classify assumptions by what their failure would do to the proposed direction.
Keep fast feedback near edits and place expensive checks at explicit later gates.
Reuse a versioned agent definition while giving each execution its own run identity.
Use ordinary code for steps fully determined by their inputs and retain model judgment for the ambiguous remainder.
Record the assumptions behind a search exclusion and retain the candidate so it can be reopened when evidence or prerequisites change.
Verify that a behavioral test fails for the intended missing behavior before implementing the smallest change that passes it.
Make uncertain requirements and competing interpretations explicit before an irreversible choice.
Translate vague actions into observable success criteria and verify each meaningful step.
Classify redesign intent and inventory nonvisual contracts before changing presentation.
Extract an explicit visual contract before coding from a reference artifact.
Use expand/migrate/contract stages when a broad mechanical change cannot land in independent vertical slices.
Choose the prototype artifact by the uncertainty it must resolve and expose relevant state.
Ask only decisions whose prerequisites have already settled, then recompute the next round.
Ask the requester who holds missing knowledge and what decision needs it, then target the actual owner.
Preserve intervening human edits while materializing agreed work incrementally.
Maintain shared rules once and store page deviations separately, with explicit override, inheritance and missing-file handling.
Treat a retrieved UI suggestion as provisional until its scope and identity fit the request.
Record an authorized implementation decision, its evidence and cost if wrong, then include it in the final handoff.
Ask one focused follow-up at a time until the current topic stops producing new information.
Keep the completion observer alive while its producer finalizes pending work.
Remove provably ineligible candidates before spending on ranking or model enrichment.
Keep the identity of the tested artifact unchanged through staging, publication and promotion, and read it back.
Check existing local and external solutions before writing a new utility, then explain the chosen reuse strategy.
Identify automated output events and exclude them before they recreate the workflow that produced them.
Build one small end-to-end user path from the existing specification before expanding all layers.
Release only temporary state created and still owned by the current operation.