Decision to delivery

Make the next useful outcome explicit, route uncertainty to the action that can resolve it, and complete the smallest artifact that serves that outcome. Treat an apparent workflow problem as a hypothesis, not a diagnosis of the user.

Establish the working agreement

Read the actual brief and current artifact before proposing more work. Inspect relevant prior dialogue as context, never as fresh authorization. Identify separately:

• the user's decision (for example, whether a role fits);
• the required deliverable (for example, a case response);
• external validation (for example, willingness to pay);
• optional reusable tooling.

Use a short round of high-value questions to establish the immediate objective, desired outcome, constraints, deadline, time budget and outstanding commitments. Check dated communications where authorized: a self-imposed deadline and a promise already made may differ. Surface that difference without silently changing either.

Propose a concrete flow with inputs, outputs, owners, time limits, stopping and reopening criteria. If the user requires flow approval, obtain explicit approval before creating a skill or applying the flow. Preserve that boundary across interruptions. Don't add an approval gate when the current request already authorizes the work.

Map and route

Map each required output to existing material, its missing consequential decision and the smallest next action. Keep case requirements distinct from attractive demonstrations or optional infrastructure. Avoid maintaining a second competing version of the argument.

For each consequential unknown, choose the route that can answer it:

Research cannot establish a person's commitment. A working demo cannot establish customer demand. A sourced observation is not proof of commercial fit. Keep these distinctions in the deliverable itself.

Record only material decisions: question, current position, source/assumption, next action, owner, stop condition and reopen condition. Keep unresolved commitments visible without generating an exhaustive new queue.

Deliver within the agreement

Prioritize complete coverage and one reliable user journey before deep polish. Reuse the existing application/data when useful; inspect and back up touched files, preserve concurrent work and review decisions, and treat synced references as read-only. Prefer an additive presentation view when it avoids changing an active research system.

Review evidence in order of its ability to change the decision or support what will be shown. Don't clear a large backlog or expand a batch merely to make a presentation appear complete. Label target candidates, hypotheses, estimates, precomputed AI output and simulated workflows accurately. Never substitute fictional records for promised real examples.

Use editable assumptions where information is unavailable and the brief permits assumptions. Show their effect and provenance. Propose future tests rather than claiming external validation. When the user rules out new input, preserve gaps instead of inventing answers or contacting people.

Bound tooling by the next learning or delivery step. Ask what outcome would justify expansion; if none would change the next action, park the enhancement. If time tightens, cut optional interactions and polish before required coverage, usable sales material, provenance or reliable navigation.

Close and reopen

Verify the artifact as its intended user will experience it: rendered behavior, the core journey, calculations, evidence links and labels. Distinguish technical completion, external validation and the user's review.

Handover: deliverable, changes, verification, material limitations, decisions made, unresolved commitments with owners, and the next action. Notify through the user's preferred review channel when available.

Stop when the agreed required outputs exist and work, evidence/assumptions are legible, and remaining unknowns have a route and owner. Reopen only for contradictory evidence, a material constraint change, a failed journey or explicit feedback. New cosmetic ideas need not reopen a completed scope.

Sending, publishing, deployment and other consequential external actions require their own appropriate authorization; approval of this operating flow does not confer it.