gap-scan

Scan a codebase, live app, or site and find the features that obviously should exist but don't — missing exports, bulk actions, search/filter, undo, history, notifications, dead-end workflows, data co

SKILL.md

GitHub

Gap-Scan

Find the features a product is obviously missing. Not what's broken (code-audit), not what's confusing (ux-audit), not hidden value in what exists — what's absent that users of this kind of product would reasonably expect.

The output is a punch list a client or team can act on: each gap cites the evidence that proves it, ranked by user value × effort, capped at 10. The report's credibility comes from what it refuses to include — no speculative wishlists, no "you could add AI."

Modes

Ask which mode if not obvious from context; default to QA mode for the user's own repo, PROSPECT mode when the target is a lead's site/app. If a gap_packet.json already exists for this target, default to RECHECK instead of re-running a full scan.

  • PROSPECT mode — fast pass on a lead's app or site. Goal: proposal ammunition. Top 5–8 gaps, each phrased as a client-facing line item ("Your customers can't reorder past purchases — every competitor offers this"). Lighter evidence bar (observed from the outside is fine), but still cite what you observed.
  • QA mode — deep pass on a repo you have code access to. Goal: roadmap input. Full inventory, file-path evidence required for every finding, up to 10 gaps.
  • RECHECK mode — re-run against an existing packet. See "Recheck mode" below.

Process

1. Inventory once, reason many times

Do ONE scan of the target and build a capability map. Do not re-crawl files per lens — every lens below reasons against the map, not the repo.

For a repo, the map covers:

  • Routes/pages and what each lets the user do
  • Models/entities and their fields (especially fields written but never read/displayed)
  • CRUD coverage per entity: create / read / update / delete / duplicate / archive / export / import / bulk
  • Endpoints with no consuming UI; UI elements with no working backend
  • TODOs, commented-out blocks, disabled buttons, feature flags that are off, dead routes
  • Background jobs, events, and data being collected

For a live app (no code), the map comes from walking the UI: every screen, every action available per entity, every dead end.

2. Run the gap lenses against the map

Each lens is a question asked of the map — cheap reasoning, no new scanning:

  1. CRUD symmetry — user can create but not duplicate/archive/export/bulk-edit? Lists with no search, filter, or sort where the data volume clearly demands it?
  2. Dark data — fields or events collected but never surfaced anywhere. (Timestamps stored but no "last updated" shown; analytics events fired into nothing.)
  3. Orphans — backend capability with no UI; UI promising something the backend can't do; endpoints nothing calls.
  4. Half-built — TODOs, disabled controls, dead-end screens, flags stuck off. Features that stopped partway are the cheapest gaps to close.
  5. Table stakes — what does this category of product normally have that this one lacks? (A CRM without activity history; an e-commerce app without order status; a SaaS without password reset.) Name the category explicitly before applying this lens, so the comparison set is honest.
  6. Workflow dead ends — a user completes step N and there's no step N+1 where one obviously belongs (created an invoice, can't send it; finished a report, can't share it).

3. Evidence gate

Every finding must cite proof or it's cut:

  • QA mode: file path(s), route, or model field that demonstrates the absence ("orders.ts has create/get/update — no export or bulk endpoint exists; OrdersList.tsx renders single-row actions only")
  • PROSPECT mode: what was observed ("Products page lists 200+ items, no search input present")

Anchor evidence to symbols, never to line numbers. Cite src/api/orders.ts → createOrder or OrdersList.tsx → <RowActions>, not orders.ts:142. Lines move on every unrelated edit; symbols don't. A recheck can then tell "the symbol moved" (finding still valid) from "the symbol is gone" (finding stale) — with line numbers both look identical.

"Users probably want X" with no anchor in the map is not a finding. The table-stakes lens still needs an anchor: name the comparable category norm AND confirm the map shows the absence.

4. Rank and cap

Score each surviving gap on two axes, no fake precision:

  • User value: High / Medium / Low — how often would real users hit this absence?
  • Effort: S / M / L — S if the surrounding infrastructure already exists (single-record action → bulk is usually S; a whole notification system is L)

Sort: High-value + S-effort first. Hard cap: 10 findings (5–8 in PROSPECT mode). Everything else gets one line in an appendix or gets cut. Resist padding — a 6-gap report that's all real beats a 10-gap report with filler.

Report structure

ALWAYS use this shape:

# Gap-Scan: [target]
Mode: [PROSPECT | QA] · Category baseline: [what kind of product this is]

## Top gaps
For each (max 10):
### N. [Gap name]
- What's missing: [one sentence]
- Evidence: [file → symbol anchors / observed behavior — never line numbers]
- Why it matters: [one sentence, user-anchored]
- Effort: [S/M/L] — [one line on why, citing existing infra if S]
- Suggested MVP: [smallest version that closes the gap]

## Appendix: minor gaps (one line each)

## Out of scope (handoffs)
[Anything found that belongs to code-audit / ux-audit / conversion-audit — one line each, don't elaborate]

## Start here
[ONE gap, by name. The first concrete action to take on it. One line on why it beats the others.]

## Start here is required and must be the last section — after the handoffs, so the report ends pointing at the reader, not at another skill.

Rules for it:

  • Name exactly ONE gap. Not a ranked shortlist, not "start with 1 or 3" — one.
  • Give the first concrete action, not the whole MVP. "Add a GET /orders/export route next to the existing getOrders handler" — something that can be started in the next 10 minutes.
  • One line on why this one first. Usually: highest value at S effort, or it unblocks others.
  • Pick it even if the top-ranked gap isn't the right first move. Ranking is by value×effort; the first move can differ (a Medium-value S-effort gap that takes 20 minutes may beat the High-value L-effort one as a starting point). If they differ, say so in the why-line.

In PROSPECT mode, add a ## Proposal framing section — each top gap restated as a client-facing benefit line — placed before ## Start here, which stays last.

gap_packet output

Alongside the report, always emit gap_packet.json so other skills and swarm agents can consume the results. The packet is a work list, not a snapshot — it carries status so "what's left?" is answerable by reading the file, never by re-investigating the repo.

{
  "packet_type": "gap_packet",
  "version": 2,
  "target": "repo path or URL",
  "mode": "prospect | qa",
  "category_baseline": "e.g. local-service business site",
  "scanned_at": "ISO date",
  "last_rechecked_at": "ISO date or null",
  "start_here": "id of the gap named in the report's Start here section",
  "gaps": [
    {
      "id": "bulk-actions-contacts",
      "name": "Bulk actions on contacts",
      "lens": "crud_symmetry",
      "value": "high",
      "effort": "S",
      "evidence": ["src/api/contacts.ts → listContacts", "src/components/ContactList.tsx → <ContactRow>"],
      "mvp": "Multi-select + bulk delete/tag on the contacts list",
      "status": "open",
      "first_seen_at": "ISO date",
      "closed_at": null,
      "closed_by": null,
      "supersedes": null,
      "superseded_reason": null
    }
  ],
  "handoffs": [{"skill": "ux-audit", "note": "checkout flow confusing"}]
}

Field rules:

  • id — stable kebab-case slug. Never reuse or renumber; a recheck matches gaps by id.
  • status — one of:
    • open — still missing, still valid
    • closed — the feature now exists (gap was actually filled)
    • superseded — the finding no longer holds for a reason other than being built: the evidence moved, the surrounding code was rewritten, or the premise turned out to be wrong. superseded is not a soft closed — nothing shipped.
  • closed_at / closed_by — ISO date, and "recheck" | "user" | "<skill or agent name>". Set on both closed and superseded; null while open.
  • supersedes — when a recheck replaces a stale finding with a corrected one, the new gap points at the old gap's id.
  • superseded_reason — required whenever status is superseded, null otherwise. One or two sentences on why the finding no longer holds: which anchor moved, what was rewritten, or — for a retraction — what the premise got wrong and what evidence disproved it. A superseded gap with no reason is indistinguishable from a finding that quietly vanished, which is the exact failure this status exists to prevent.
  • Never delete a gap from the array. Status transitions only — the history is the point.

Recheck mode

Findings rot. Between a scan and the next look at it, code gets rewritten, endpoints move, and gaps get filled without anyone updating the packet — so a stale packet quietly reports things that are no longer true while missing gaps introduced since. Recheck is the fix, and it is not a re-scan: it starts from the packet.

Trigger it when a gap_packet.json exists for the target and the user asks "what's left", "is this still accurate", "recheck", or re-runs gap-scan on the same target.

1. Load, don't re-derive

Read gap_packet.json first. Every open gap is a claim to be tested. Do not start a fresh inventory — that throws away the status history and produces new ids for old gaps.

Answering "what's left to do?" alone stops here: filter status == "open" and report. That's a file read, not an investigation. Only continue to step 2 if the user wants accuracy re-verified.

2. Re-verify each open gap's evidence

For each open gap, check its evidence anchors still exist and still demonstrate absence:

  • Symbol present, gap still absent → stays open.
  • Symbol present, feature now exists → closed, closed_by: "recheck".
  • Symbol gone / file gone / surrounding code rewritten → the finding can no longer be verified. Mark superseded. If the gap still holds under the new structure, add a NEW gap with a new id and supersedes pointing at the old one.
  • Premise was wrong (the capability existed elsewhere all along, or the evidence was misread) → superseded, and say so plainly in the report. A wrong finding retracted is worth more than a wrong finding left standing.

Every transition out of open writes superseded_reason (for supersedes) and closed_by. If the fix was attempted and then reverted, say that — "implemented, produced N false positives and zero true findings, reverted" is the most useful sentence a recheck can leave behind.

3. Scan for gaps introduced since

Only after step 2. Run the lenses (§2) against what changed since scanned_at — new routes, new models, new half-built work. New findings get new ids, status: "open", first_seen_at set to today.

4. Report the deltas, not the whole list

# Gap-Scan Recheck: [target]
Last scanned: [date] · Rechecked: [date]

## Still open (N)
[one line each — name + why still valid]

## Closed since last scan (N)
[one line each — what was built]

## Superseded (N)
[one line each — why the finding no longer holds, and whether a replacement gap was filed]

## New gaps found (N)
[full finding format for each]

## Start here
[same rules as a fresh scan — ONE gap, first action, why]

Bump last_rechecked_at. Write the packet back — recheck always persists, or the next recheck starts from stale data again.

Boundaries — what gap-scan never does

  • Never reports something that exists but is broken, slow, or insecure → code-audit
  • Never critiques the usability of what exists → ux-audit
  • Never evaluates persuasion or conversion of existing pages → conversion-audit
  • Never proposes productizing existing internal assets → that's opportunity-mining (WWBD), a different question
  • Never recommends features with no evidence anchor ("add AI", "add a mobile app") — if the map doesn't imply it, it doesn't ship in the report
  • Never answers "what's left?" by re-investigating the repo when a packet exists → read the packet's status fields
  • Never ends a report without ## Start here — a list with no first move is an unfinished report