Skip to content
Describes how WilsonOS is designed to work.What works today →

End to end

Two illustrative walkthroughs, told as one story each. The other examples show single workflows with the real outputs; these two show how the steps join up. The first starts from nothing and ends with a product on sale; the second starts from a book that already exists and ends with the next experiment, not a rewrite. Each numbered item names the step of How it works it belongs to. Whether each step is available depends on installed capabilities, connected services, permissions and verified implementation; see What works today.

This is an illustrative end-to-end workflow. It shows the five steps and the shared business records fitting together when a real outcome runs through them. Each numbered item below names the step it belongs to. Availability of each step depends on installed capabilities, connected services, permissions, and verified implementation.

The setting. Code Pineapple is a children’s book imprint. Its brand agent is Scout. The owner has noticed parents asking for something to keep six-to-ten-year-olds busy over a school holiday and wonders whether an activity-and-puzzle resource would sell.

  1. Research the opportunity — step 2 · talk to your agent

    The owner opens a conversation with Scout: “Is there a real audience for a holiday activity book, and what would it need to be?” Scout already has Code Pineapple’s positioning, past titles, and customer language in context. It runs a research flow against authorised sources, and comes back with findings, the assumptions it made, and a proposed business goal. The owner corrects one assumption in chat; that correction is recorded as a decision, not lost in the thread.

  2. Shape the product and offer — step 2

    Together they define the resource (a printable activity pack), the buyer (a parent, not the child), the promise, the price point, and the acceptance criteria — what “done and good” means before anything is made. A product record and an offer record now exist and link to each other.

  3. Plan the work — step 3 · it does the work

    Scout creates tasks under a project, selects the flow that fits a printable product launch, and lays out milestones. The plan names the points where a human decision is required (cover approval, price confirmation, publish). The owner approves the plan; it becomes version zero of the task record.

  4. Create the deliverables — step 3

    Agents write, design, and assemble: the activity pages, a cover, a sales page draft. Each piece is produced by a named skill in a scoped working environment, and each run records what was invoked with what inputs. Image generation goes through the one media gateway, so every image has a receipt and a cost.

  5. Review and approve — step 4 · you review and publish

    The owner opens the task in the Workbench. Deliverables appear as cards; the cover opens in the same viewer as the sales page and the PDF. Feedback attaches to the specific version it was given on — “warmer palette on v2” — and Scout produces v3 without losing v2’s history. Approval is on a version, not on a name.

  6. Publish and distribute — step 4

    The approved pack and sales page are published through supported connections. The publication record captures the external identifiers that came back — the listing, the page — so that results can be matched to what was actually shipped.

  7. Measure the outcome — step 5 · you measure

    Visits, inquiries, sales, cost, and feedback appear with their source and freshness visible. The dashboard says what it can prove (“forty-one visits from the sales page, six purchases”) and what it cannot (“attribution to the original research assumption is not observable”).

  8. Improve deliberately — step 5 · it learns

    Scout proposes one product change and one capability change, each with the evidence behind it. The capability change — a tweak to the cover-design skill — is tested against the current version before anyone promotes it. The accepted version and the reasoning are kept; see The improvement contract.

flowchart LR
  AU[Audience need] --> P[Product] --> O[Offer] --> T[Task] --> R[Run] --> V[Artifact version] --> PB[Publication] --> RS[Results]
  RS -. next proposal .-> T

audience → product → offer → task → run → artifact version → publication → results. The model preserves these links where the underlying systems provide evidence, and states the gap where they do not.

  • The owner never chose a tool. They stated an intent, answered questions, and approved versions. Scout chose the flow and the skills, and said which ones it chose.
  • Every “yes” was on a specific thing. A version, a price, a plan — never “the cover” in the abstract.
  • The loop closed. Results fed the next proposal. Nothing was rewritten because one number looked good.

The first example started from nothing. This one starts from a product that already exists and shows the five steps applied to marketing rather than creation — and, at the end, what “improve” does and does not mean.

The setting. RiseBlox publishes technology and education titles. Its brand agent is Prime. One book has steady but flat sales, and the owner wants more parents to discover it.

  1. Choose the goal — step 2 · talk to your agent

    “Help more parents discover this book.” The owner selects the product and, with Prime, defines what success looks like for this campaign: a number, a window, and the funnel event that counts.

  2. Develop the strategy — step 2

    Prime uses RiseBlox’s brand context and prior research to propose the audience, the message, the offer framing, and the channel — Pinterest, because that is where this audience plans and saves. The strategy is a reviewable document, and the owner’s edits become decisions carried forward.

  3. Create the Pinterest posts — step 3 · it does the work

    A selected flow produces the copy and the visuals for a set of pins. Each pin keeps its attribution: which task, which run, which skill and version. Image generation runs through the media gateway with a cost receipt per image.

  4. Review and schedule — step 4 · you review and publish

    The owner approves specific asset versions in the Workbench. Calendar entries are created for the planned posts; each entry refers to its post and its campaign rather than copying either.

  5. Publish through the connection — step 4

    The authorised Pinterest connection publishes on schedule. Success or failure is recorded per post, along with the external post identifier — the thing that will later let a click be matched to a pin.

  6. Collect available signals — step 5 · you measure

    Impressions, clicks, and downstream outcomes appear where tracking exists. Where it does not — a purchase on a third-party store that cannot be linked to a pin — the dashboard says so rather than inventing a number. Attribution limits stay visible.

  7. Compare performance and feedback — step 5

    Two things are compared, and kept apart: the owner’s visual approval of the pins, and their business performance. A pin the owner liked least may have performed best. Audience, offer, and timing are examined alongside the creative, because the creative is rarely the only variable.

  8. Run the next experiment — step 5 · it learns

    Prime suggests and tests changes: a different hook, a different posting time, a different offer framing. What it does not do is rewrite the shared Pinterest skill because one post performed well. A proposal with evidence goes through the improvement contract; the skill changes only if the variant beats the current version on explicit criteria.

The same pin appears in every area of the workspace (see How it works), and in none of them is it copied:

Area What it sees
Work & deliverables The task that created it and the approved version
Conversations The chat where the goal was set, with a smart link to the task
Capabilities & automation The flow and skill that produced it; the schedule that published it
Business performance Its impressions, clicks, and the outcomes that can be attributed
Learning & improvement The proposal it motivated and the test that followed

And the shared records tie them: the product identifies what it promotes; the task tracks creation; the calendar schedules it; the connection publishes it; performance measures it; learning proposes improvements.

  • Approval and effectiveness were measured separately. The system is built so that “I like it” and “it worked” are two different columns.
  • Attribution limits were stated, not hidden. A stage in the funnel is a definition; only observed events with a source count as results.
  • The skill survived a good week. Improvement is a proposal with evidence, not a reaction to one number.

Sources: input/WilsonOS_Overview.html §04 · Last verified 2026-10-06