| name | requirement-gap-analysis |
| description | Interrogate a pre-design requirements doc, stakeholder brief, epic, or freeform notes for missing information โ before any solution has been proposed. Use whenever the user asks what's missing from requirements, wants help scoping a feature before writing an RFC or design doc, shares a rough feature idea and asks what to consider, or asks "what should I ask before we design/build this". |
requirement-gap-analysis
Interrogates a pre-solution input โ a raw requirements doc, stakeholder
brief, Jira epic description, meeting notes, or freeform text describing a
problem/feature request โ for what's not yet known that needs to be
known before anyone should start designing a solution. This runs one
stage earlier than rfc-review: rfc-review judges whether a proposed
design holds up (including a "completeness" check against what the RFC
itself claims to cover); this skill has no proposal to judge at all โ it
asks category-based questions the input document doesn't even attempt to
answer, because it isn't structured as a design doc yet.
No script layer โ this skill is judgment only.
Step 1 โ Resolve the input
Accept a file path or pasted text only โ don't fetch external doc-platform
URLs (they return login walls); ask for pasted content or an export
instead. If nothing is given, ask what to analyze.
No template expected. Don't penalize the input for lacking RFC-style
structure โ bullet points, a paragraph, a transcript excerpt, and a
formal brief are all equally valid input shapes here.
Step 2 โ Sanity-check the input's stage
Before analyzing: if the input already reads as a proposed solution
(has a Design/Approach/Solution section, describes a specific technical
implementation rather than a problem to solve), this input is past this
skill's intended stage. Say so plainly and suggest rfc-review instead โ
don't silently run gap analysis on a doc that's actually ready for design
review; the two skills would produce overlapping, confusing output on the
same input.
Step 3 โ Interrogate across categories
Work through each category, skipping one only if genuinely inapplicable
to this input (say why):
- Functional โ unstated behavior, missing use cases, unclear scope
boundaries (what's explicitly in/out of scope, if not stated).
- Non-functional โ performance, scale, availability, latency
expectations never stated.
- Edge cases โ error states, concurrent access, empty/boundary
conditions not addressed.
- Stakeholders โ who's affected or should be consulted that isn't
mentioned (legal, security, other teams whose systems this touches,
end users with accessibility needs).
- Data โ what data is involved, where it comes from, ownership/
retention/privacy implications.
- Compliance/regulatory โ anything industry- or jurisdiction-specific
left unaddressed, when inferable from context (don't invent regulatory
concerns with no basis in the input's domain).
Codebase grounding (only when run inside a repo, same convention as
rfc-review): for concretely checkable claims about the existing system,
grep for them, and label findings verified/unverified accordingly.
Don't attempt exhaustive verification โ only what's cheap and unambiguous.
Step 4 โ Assemble findings
Each finding: category, the specific question that needs answering
(not "X is missing" โ the actual question someone should go ask), and why
it matters (what gets designed wrong, or costs more to fix later, if this
stays unanswered). Tag each with a severity tier, reusing rfc-review's
scale directly:
- Blocking โ can't responsibly start designing without this answered.
- Should-Fix โ meaningfully de-risks the design if answered now, not
fatal to skip.
- Nice-to-Have โ worth asking, low stakes either way.
- Question โ genuinely open, more a clarifying question than a gap.
No confidence axis. Unlike security-review/performance-review's
judgment calls about whether code is actually buggy, a gap is a gap by
construction โ the input simply doesn't address it. There's no "might be
a false positive" dimension to a missing answer.
Step 5 โ Write the report
Write GAP_ANALYSIS.md next to the input file (or cwd if piped/pasted),
and show it inline โ both. Structure: findings grouped by category, each
tagged with severity, containing the question and why it matters. No
overall verdict โ there's no proposal to approve/reject, only questions
to go answer.
Every invocation is a fresh analysis โ this skill does not track prior
analyses or diff against an earlier version of the same input in v1.
Step 6 โ Offer to resolve the gaps
After presenting the report, offer via AskUserQuestion to walk the user
through the Blocking questions one at a time โ one question per turn,
following the same convention as this repo's interactive skills โ
recording each answer, then append an ## Answers section to
GAP_ANALYSIS.md capturing them. This turns the gap list into
design-ready input. If the user declines, stop after the report.
Step 7 โ Composition with rfc-review
When rfc-review later runs on the RFC that eventually gets written from
this input, and finds a sibling GAP_ANALYSIS.md, it should note under
its own Completeness dimension which flagged gaps got addressed in the
RFC and which didn't โ same explicit sibling-check convention as the rest
of the pipeline, not a hard dependency.
Rules
- Don't apply this skill to a doc that already proposes a solution โ flag
it and point to
rfc-review instead.
- Don't invent compliance/regulatory concerns with no basis in the
input's actual domain.
- Don't attach a confidence score to findings โ a gap is a gap by
construction.
- Don't invent an overall approve/reject verdict.
