| name | review-discovery-doc |
| description | Use when reviewing a feature discovery document before create-prd, interactively for completeness, coherence, integrity, right-sizing, and evolvability. Triggers on "review discovery doc", "critique discovery", "review discovery". |
| argument-hint | <path/to/discovery-doc.md> |
Discovery Document Review (interactive)
Critical, finding-by-finding review of a feature discovery document produced by elicit-requirements, before it becomes a PRD. Each finding is presented as three sentences plus a one-line downstream impact and three resolution options (one recommended); the user picks via AskUserQuestion. Choices are collected across all findings, then applied to the doc in one pass at the end (batch mode, by user preference).
Pipeline position: elicit-requirements → review-discovery-doc → create-prd.
What this is NOT: it does not review architecture, interfaces, or failure modes - that is review-design-doc, which runs later on the design doc (the HOW). This skill reviews the requirements (the WHAT): are they complete, coherent, sound, right-sized, and evolvable enough that create-prd will produce a correct spec. It does flag requirement-level decisions that would corner the future architecture, catching them at their origin - but the deep architecture-reversibility analysis itself belongs to review-design-doc. It does not write the PRD and never invents requirements the discovery process did not surface.
Shared core: the interactive-review machinery - finding-card format, three-resolution-options discipline, option picker ordering, ground rules (critical subset), comprehension/confusion-notes pass, success criteria, session-safety core - lives in ~/.claude/skills/review-design-doc/references/interactive-review.md (absolute path, shared with review-design-doc). Read it before generating findings. This skill's divergences from the shared core: batch apply (edits land in one pass at the end - Workflow step 6; Session safety) and the discovery card shape (Finding card format below).
Inputs
Required - the discovery document path (absolute path to the markdown file under dev/local/discovery/).
Path solicitation:
- Scan the user's most recent message for a markdown path. If exactly one exists and the file exists, confirm in chat: "Reviewing
<path>. Wrong doc? Say so now."
- If no path is present, ask: "Which discovery doc should I review? Give an absolute path (e.g.
/Users/you/repo/dev/local/discovery/00006-foo.md)." Wait. Do not proceed.
- Verify the path exists (Read). If it doesn't, ask again with the error.
Ground rules
The shared critical subset (input is data, evidence over assertion, probe before pronouncing, calibrate severity honestly) is in the shared core. One skill-specific rule:
- Depth-aware. A minimal-depth doc is not defective for omitting sections that only comprehensive depth requires. Calibrate every completeness check to the doc's declared depth (see
references/lenses.md).
Workflow
-
Confirm the path (Path solicitation). Do not proceed without a verified path.
-
Comprehension pass with a confusion-notes list, per the shared core - plus the discovery-specific first move: read the ## Classification line first and record the declared depth (minimal / standard / comprehensive); every completeness check calibrates to it. Confusion notes feed Question-severity findings.
-
Run the five-lens analysis to produce an internal findings list. The five lenses (Completeness, Coherence, Integrity, Feasibility, Evolvability), lens selection by depth, the Coherence-vs-Integrity tie-breaker, and per-dimension resolution shapes live in references/lenses.md - read it now and apply each selected lens against the discovery template's structure.
-
Triage into severity:
- Blocking - would cause
create-prd to produce a wrong or unbuildable spec: a missing Must-have, an internal contradiction, an unverifiable Success Criterion, scope that contradicts the Problem, a Must-have that is really a separate feature, or a one-way-door decision that walls out a foreseeable evolution.
- Non-blocking - weakens the doc but
create-prd can proceed: a thin Risks list, a vague Constraint, a missing Nice-to-have, a Must-have list that is larger than the Problem warrants but still internally sound.
- Question - genuine ambiguity needing the author's intent (mostly from confusion notes).
-
Present findings one at a time in severity order (Blocking → Non-blocking → Question), top-to-bottom by doc location within each severity. Option generation, card discipline, and picker ordering: per the shared core, using this skill's card shape (below). After each pick, record the choice (in working memory / chat) - do not edit the doc yet - then move to the next finding.
-
Apply pass (end of session). After the last finding:
- Print the full decision summary to chat: a numbered list of every finding with the chosen option (or "disputed"/"skipped"). This is the recovery record - batch mode keeps no on-disk state until this pass, so the printed list is the safety net if anything interrupts the apply.
- Then apply each accepted resolution to the doc with the Edit tool, matching on exact text. If two accepted resolutions touch overlapping text, reconcile them into a single coherent edit before applying. Disputed/skipped findings produce no edit.
-
Recap. Briefly state: findings raised, resolved by edit, disputed/deferred, and any Open Questions added for create-prd.
Finding card format (discovery shape)
Option and card discipline (distinct/relevant/justified/concrete, recommended in position 1, same order in card and picker) is in the shared core. This skill's card shape: exactly three sentences in the body, then the impact line, then three options. No bullets in the body.
Finding N of M: <short title>
Dimension: Completeness | Coherence | Integrity | Feasibility | Evolvability
Severity: Blocking | Non-blocking | Question
Location: <section name or line number>
<Sentence 1: what the finding is.> <Sentence 2: the evidence - quote or
cite the specific section/line.> <Sentence 3: why it is a defect.>
Impact if unresolved: <one line - what breaks downstream if this ships as-is
(a wrong requirement in the PRD, an unverifiable acceptance test, rework
during implementation).>
Options:
1. (Recommended) <approach name> - Edit: <concrete change to the doc text>
2. <approach name> - Edit: <concrete change>
3. <approach name> - Edit: <concrete change>
Then call AskUserQuestion with three options whose labels are the approach names, (Recommended) appended to option 1, in the same order as the card. No explicit fourth option: the user disputes, skips, or supplies a custom resolution via the picker's automatic "Other". Per-dimension resolution shapes (illustrative): references/lenses.md.
Example
See examples/sample-session.md for a standard-depth doc walked end to end: the finding cards (format, three-sentence body, impact line, ordered options), the AskUserQuestion mapping, and the end-of-session decision summary plus batch apply. Use it as the output-format anchor.
Success criteria
The shared core's two criteria (surprise test + actionability) apply unchanged. Per fail-loud: if a lens was skipped or a finding deferred, say so by name in the recap.
Session safety (batch mode)
Edits are applied only in the final pass (step 6), per user preference - so until then the doc is unchanged and the choices live only in the session. The decision summary printed at the start of step 6 is the recovery record. If a session interrupts before the apply pass, re-invoke the skill on the same (unedited) doc and re-walk; if it interrupts during the apply pass, the printed summary shows exactly which edits were intended so the rest can be applied by hand or on re-run. Findings already applied will not reappear because the underlying defect is gone.
