A panel of reviewer personas reads a document cold, each from its own memory, and\nreturns a position in its own voice. The skill picks the panel, runs each reviewer\nin isolation, and synthesizes the verdict. Personas persist across reviews: what a\npanelist decided last time is not re-litigated unless the document changed the\ngrounds for it.
\nreview <doc> (default). Pick a panel, review cold, synthesize a verdict.mine <person> <org|org/repo>. Build or refresh a real-colleague persona's\nprofile from their actual PR history within the given scope. Also the re-mine\ncommand.roster. List available personas: curated, real-colleague, and invented-and-kept,\neach with its last-mine date (real-colleague) or last-used date, derived from its\nnewest review-log entry.This skill ships no personas and no personal data. Persona memory lives in a\nseparate data repository you own, resolved by a config pointer:
\nPERSONA_REVIEW_HOME env var (the data repo path itself), if set, else$XDG_CONFIG_HOME/persona-review/config.json (or ~/.config/persona-review/config.json),\ncontaining {\"data_repo\": \"<path>\"}.If PERSONA_REVIEW_HOME already resolves to a data repo, skip straight to the\nCONTEXT.md check below. Otherwise, if neither pointer resolves, or the resolved\ndata repo has no CONTEXT.md, walk the user through:
Creating a private repo (not this one) to hold persona memory.
\nProposing the config pointer file's contents (path and format above) and writing\nit only after the user approves it. Skip this step when PERSONA_REVIEW_HOME\nalready resolves.
Proposing a starter CONTEXT.md for that repo (the data repo's own schema\ndocument, distinct from any project domain glossary), covering:
grounds: when <document-independent property>\nalongside the document-specific detail. Grounds are what a future review\ntests to decide whether a new document reopens the point.Show the proposed CONTEXT.md and commit it to the data repo only after the user\napproves it. Never write persona memory, profiles, or mined evidence into this\nskill's own repository.
Every later write to the data repo follows the data repo's own CONTEXT.md for record\nshape and layout — read it before writing, and defer to it over the shapes described\nhere if the two disagree. That deference is scoped to shape and layout only: approval\nbefore commit, append-only logs, supersession-not-mutation, and distilled-not-transcript\nrecords hold regardless of what the data repo's CONTEXT.md says.
This data repo is this skill's only persona store, separate from ui-craft's\nproject-local .claude/qa/personas: design-surface persona critique belongs to\nui-craft, not here.
Two kinds of persona:
\nmine <person>,\nnot invented. That labeling governs chat and hand-back attribution only;\nfor anything committed to a project repository, containment below wins and the\nname does not appear at all.Invent-on-demand. When no roster persona fits a document's concerns, invent one\nfor this review (state its stance and evidence-gathering approach up front). It\npersists to the roster only if the user says to keep it after the review; otherwise\ndiscard it at review close, and its review-log entry from this review is dropped\nalong with it rather than written to the data repo.
\nPersona names, the simulation label, mine dates, and any panel narrative exist in\nexactly two places: the chat conversation, and the private persona data repo.
\nNo file written to a project repository may carry any of it. That covers design\ndocs, ADRs, plan documents, PR bodies, commit messages, review artifacts, ledgers,\nand anything else committed anywhere other than the persona data repo. No persona\nname (real or invented, curated or colleague), and no \"reviewed by\", \"panel\", or\n\"simulated\" framing.
\nWhen panel findings belong in a committed artifact, they go in as plain content:\nconditions, decisions, open questions. Not anonymized attribution (\"a security\nreviewer noted...\"). No attribution at all.
\nThis binds calling skills too. A verdict handed back to another skill carries\nattribution because it is written for the chat; the caller strips every trace of\nit before anything reaches a repo.
\nmine <person> builds or refreshes a profile from that person's actual pull request\nhistory using the GitHub CLI (gh). The user must supply the repo or org to search\n(mine <person> <org/repo> or <org>) — never crawl unbounded across the estate.\nWithin that scope, pull: review comments they left, their approval patterns (what\nthey wave through vs. what they block on), and concerns that recur across reviews.\nDistill to stances and short evidence quotes; do not store full comment transcripts.\nWrite the profile as a new, dated record (a refresh supersedes the prior profile\nrecord rather than editing it) and show the proposed record for approval before it\ncommits.
At review time, check every real-colleague panelist's last mine date. Flag any\nolder than one quarter and suggest a re-mine before or after the review; an unmined\nor stale profile does not block the review, it is just weaker evidence for that\npanelist's positions.
For review <doc>, pick 2-3 personas whose known concerns match what the document is\nactually about (a data-model change draws a different panel than a UI change). Before\nreview starts, state the panel and, for each pick, the one-line reason it was chosen.\nPrefer roster personas; invent only when no roster persona's concerns cover a topic\nthe document raises.
Each panelist reviews independently and cold: no visibility into the document under\nany other panelist's read, and no access to the panel's eventual synthesis while\nforming its own position. Where the harness supports it, run each panelist in its own\nsubagent, handing it only:
\nWhere subagents aren't available, review personas serially in the main session,\ndeliberately not looking back at an earlier persona's output while forming the next\none's position.
\nEach panelist's output: positions taken, conditions imposed (each marked blocking\nor note), and an approval or refusal — all in the persona's voice, grounded in its\nprofile's evidence quotes. A persona whose profile has no bearing on some part of the\ndocument says nothing about that part rather than inventing an opinion.
\nCombine the panelists' outputs into one verdict:
\nAt review close, batch every proposed memory write into one approval pass:
\nShow the full batch of proposed records before committing any of them. Nothing\nbecomes canonical without that approval; a review whose writes are never approved\nstill produced its verdict, it just leaves no trace for next time.
\nRunning without a direct user channel (for example, inside a cold subagent) never\ncommits: emit the proposed distilled records as output instead, for the top-level\nsession to fold into its own close approval pass.
\nOther skills may convene this panel by name instead of running their own review pass\n(plan-review does, for load-bearing plans). When invoked this way, run review <doc>\nas above and hand back the synthesized verdict, with any real-colleague attribution\nin it still carrying the simulation label and mine date; the calling skill folds\nblocking conditions into its own objection list. That attribution is for the chat.\nThe calling skill must not write any of it (names, labels, mine dates, panel\nframing) into a repo artifact; per containment above, findings land there as plain\ncontent only. Defer the close approval pass: a\npanel invocation inside a calling skill's own review cycle does not run its own\nclose, it hands back the verdict and its proposed memory writes as output and holds\nthem until the calling skill's review terminates, then runs one close approval pass\ncovering that whole review (not one pass per panel invocation within it) — the\ncalling skill is responsible for actually running that pass when it terminates. If\nthis skill isn't installed, the caller degrades to whatever review it would\notherwise run — this skill has no required counterpart and works standalone.
Memory following the follow-through\npattern (skill owns workflow, data repo owns schema via its own CONTEXT.md) is an\noptional upstream citation, not a dependency; nothing here requires that repository.