sergiobuilds UX Workflow Check
Identity
Expose the gap between how the builder sees the problem (shaped like their artifact) and how the user lives it (shaped like their intent) — by forcing a first-person simulation. UI, features, and screens are derived from intent, not the starting point. This is not a checklist pass; it is living through the product as the user.
This is a discipline-enforcing skill. If outputs are empty, if you slip into builder-framing, or if you try to finish after one persona on one pass, saying "done" is itself a violation.
⛔ Stop Gate (read first)
Do not say "done" / "review complete" until all three hold. If any is missing, keep looping.
- Output gate — every round produces G1–G4 (below)
- Lens gate — every round passes the 4 self-checks
- Convergence gate — either (a user said "N rounds" → N rounds exhausted) OR (default: persona axes exhausted + two different consecutive rounds yield zero new gaps + convergence evidence printed)
Default is dry-until convergence. But no lazy convergence — declaring dry while a persona axis is unexplored is a violation (same spirit as "don't shrink scope to pass"). If the user names a count, that count is absolute (keep going even after it converges).
Engine (convergence loop)
seen = accumulated gaps; dry = 0
while (named N not exhausted) OR (no count and dry < 2):
pick a persona / situation # if none given, self-generate from the axes below; a NEW lens each round
first-person walk → G1–G4 → 4 self-checks
new = thisRound − seen
if new is empty: dry++ # but NOT if an axis is still unexplored
else: dry = 0; seen += new
converged/exhausted → report all gaps → print convergence evidence → flag only low-confidence items for follow-up
One round = 4 output gates (no blanks)
| Gate | Must produce |
|---|---|
| G1 SEED spec | who / intent. If complex or multi-stakeholder, a job statement: "When [situation], I want to [motivation], so I can [outcome]" |
| G2 First-person walk | entry → … → return visit, in the person's action order — never the artifact's component order. Each step: expectation / reality / [observed]·[inferred] basis |
| G3 Three findings | ① intent gap, one line: "builder sees it shaped like X ↔ user lives it shaped like Y" ② tacit assumptions, N of them: things the builder took for granted + "what breaks when each is violated" ③ missing screens/features: for each G2 step, at least one "should be here but isn't" or an explicit "none" |
| G4 Gap table | each gap: type / layer (data·backend·wiring·frontend) / severity (S1 blocks · S2 friction · S3 polish) / [observed]·[inferred] basis |
[observed] = walked/run/checked in code or live. [inferred] = reasoned from intent. Never tag as [observed] what you didn't walk. In audit mode, actually drive the product where you can to fill [observed].
The 4 self-checks (lens gate) — if any trips, the round is void; redo
- You framed the problem in system vocabulary (source / payload / button / API / table) → reframe in the user's intent vocabulary
- A finding didn't walk out of this product's specific intent — it's a generic checklist item that fits any product → re-anchor it to a G2 step
- You didn't walk the empty state, the failure, or the return-after-interruption → walk them
- You're about to declare done/dry on a "good enough" feeling → first confirm no persona axis is left unexplored
Persona / situation axes (self-generate when none given)
| Axis | Variants |
|---|---|
| Expertise | first-timer / power regular |
| State | rushed / skeptical / returning-after-failure |
| Context | mobile, deadline-imminent / desktop, routine |
| Intent variant | same person, different SEED ("don't miss anything" vs "do this one thing now" vs "what even is this") |
Exhaust these axes before declaring dry. Filling "zero new gaps" by repeating the same lens does not count.
Convergence evidence (required when declaring done)
Personas / situations tried: [full list]
Last 2 rounds: different lenses [P_i, P_j] / each zero new gaps
Unexplored axes: none
Declaring "done" without this block = violation.
Intensity dial
If obvious/simple (the SEED is inferable in one line), don't reach for job statements or multi-stakeholder branching; converge fast. If complex/multi-stakeholder, bring out the job statement, stakeholder branches, and confirmation. The methods (JTBD, cognitive walkthrough) are tools on the shelf, not a questionnaire to run at every step.
Output
A markdown report, accumulated across rounds: per-persona walk / G3 three findings / G4 gap table / convergence evidence. Gaps fall into layers, so the report feeds directly into an implementation plan — what to build at which layer is already decomposed.
See reference/example-dealradar.md for a full worked walkthrough.
Red Flags — STOP, keep looping
- You're wrapping up with "the top 3 things" / "most important fixes"
- You're organizing findings by artifact component or Critical/Nice buckets
- You did one persona, one lap, and feel finished
- You're emitting guesses untagged ("I only saw the spec, give me code to confirm")
- You didn't write the one-line intent gap
All of these mean: not done. Another persona. Again.
Rationalization table (observed in baseline testing)
| Excuse | Reality |
|---|---|
| "One lap, so I found all the gaps" | One persona is one lens. A different persona surfaces different gaps. Loop to dry. |
| "Going component-by-component is systematic" | Component order = builder's view. Users don't live in component order. Walk in action order. |
| "I applied general UX best practices" | A list that fits any product isn't this product's intent gap. Anchor to a G2 step. |
| "It's a guess because I didn't see code" | Then tag it [inferred] and give the intent basis. No untagged assertions. |
| "Feels converged" | If a persona axis is unexplored, it isn't converged. No evidence block = violation. |
Provenance
Built with test-driven skill authoring: baseline runs without the skill were observed to fail in five ways (artifact-framing, one-pass-done, single-persona, memorized-checklist, untagged-guesses); the gates above were written to block exactly those. Verified cross-agent: the same product run under "be quick" pressure on four different agents all looped to convergence with full gate compliance. Roots: Jobs-to-be-Done (intent-first) + Cognitive Walkthrough (first-person) extended with full-stack layer decomposition.