Back to Lobehub

Trace Schema

.agents/skills/product-design/references/trace-schema.md

2.2.117.6 KB
Original Source

Trace Schema

The T of SCLPT. Without structured evidence, there is no pattern recognition — just a feeling that the session went well.

A design session leaves three artifacts. Two are for humans. The third is for the system, and it is the one that gets skipped.

ArtifactAudiencePurpose
Design specreviewersWhat we decided and why
PrototypereviewersWhat it feels like
Reality-check logthe skillWhat the business overturned — feeds the Pattern Base

The reality-check log

One row per assumption that met the business model. It lives as a section inside the spec — a separate file goes unread.

FieldMeaning
AssumptionWhat was believed, in the words it was believed in
What is trueThe business fact that overturned it, in domain language
ModelWhich of Cooper's three models this is a fact about (see layer-model.md)
Verdictoverturned | confirmed | refined
PatternP-nn if an existing pattern predicted it; NEW if it did not

Two rules keep this honest:

  • "What is true" is written in domain language, not implementation language. Not "the enum has four values" — *"the business models four kinds of agent-to-human message, and the surface renders one." If you cannot state the fact without naming a table, it is probably not a product finding at all (see layer-model.md).
  • NEW is the whole point. Every NEW row is a hole in the Pattern Base, and Step 6 closes it.

Worked rows — from the session that seeded this skill

AssumptionWhat is trueLayerVerdictPattern
The surface only ever has agent errorsThe business models four kinds of agent-to-human message — a decision to rule on, a deliverable to accept, an insight, an error. All four are produced daily; one is shownimplementationoverturnedNEW → P-04
paused means the user suspended itIt means a human is blocking the agent — pending review. It is the most urgent state on the page, not a "later" bucketimplementationoverturnedNEW → P-01
"Request changes" leaves a commentIt re-tasks the agent: resolves the item and re-runs it with the comment as inputimplementationoverturnedNEW → P-02
"Accept task" is a real actionAssignment is immediate and unilateral. There is no "offered, pending acceptance" state — the button implies you could decline, and you cannotimplementationoverturnedNEW → P-05
There is a project conceptThere is not. The section named "Project" is a knowledge libraryimplementationoverturnedNEW → P-03
We can show the team what's being discussedConversations have no notion of belonging to a member. Showing them is a domain change, not a layout choiceimplementationoverturnedNEW → P-06
A per-member unread badge is a UI decision"Unread" is a property of the conversation, not of a personimplementationoverturnedNEW → P-06
Surfacing the unmodeled work needs new plumbingThe business already models "running" and "unread" work; it was simply never asked forimplementationoverturnedNEW → P-10
Elapsed time per run can be shownThe business does not model a run's duration at allimplementationoverturnedNEW → P-11

Nine assumptions, nine overturned, all NEW — the cold-start signature. A mature run should be mostly confirmed, with one or two NEW.

Reading the log

The shape of the log is the diagnosis of the session:

ShapeWhat it means
Many overturned, all NEWCold start. The Pattern Base was blind here — harvest it.
Many overturned, all citing P-nnCheck whether the relevant patterns were read and applied; this may be a process gap.
Mostly confirmed, one or two NEWHealthy. The system is working.
Zero rowsEither a trivial surface, or — far more likely — nobody grounded anything. Treat with suspicion.
Zero NEW across a documented roundNo new gap was found in the inspected scope; check coverage before shifting effort.

If evidence keeps overturning things the Pattern Base already covers, inspect the run's pattern selection and application before adding more patterns. Do not assume the guidance itself is complete or correctly scoped.

Recording coverage

Name the sources and lifecycle areas checked, unavailable evidence, contradictions and low-confidence claims. Without this record, zero new findings is inconclusive rather than a saturation signal.

Marking debt on the prototype itself

Anything the design shows that the business cannot back gets a visible NEW tag on the mock, not just in the spec. A reviewer looking at a picture must be able to tell which parts are real.

Total time on this: 11h 06m   [NEW]   ← the business does not model run duration

This is the cheapest defense against P-05-class theatre: if you cannot name the business event behind an element, you must draw the tag.

Why the log, and not just a good memory

The Pattern Base is reviewed long-term guidance; the reality-check log is task-local evidence. Include the log in a requested spec, or summarize its material findings in a read-only response. Do not create an artifact solely to satisfy the process.