| name | rumil-review |
| description | Structured review of all research on a rumil question. Loads the subtree, all recent calls targeting the question (with brief traces), and any existing confusion-scan verdicts, then asks Claude to produce a ranked punch list of problems worth fixing — with per-item attribution and a suggested action (dispatch, apply_move, edit_prompt, or ignore). Use when the user wants to audit a question's research, not just look at individual calls. |
| argument-hint | <question_id> |
rumil-review
Loads a full review context for one question: its subtree, shape
diagnostics (structural health, rating distributions, self-reported
call signals), every recent call that targeted it (with compact trace
summaries), and any existing confusion-scan verdicts from the scan log.
Then you produce a punch list.
Context loading
setopt no_glob 2>/dev/null; set -f; PYTHONPATH=.claude/lib uv run python -m rumil_skills.gather_review_context $ARGUMENTS
Your job
Read the loaded context carefully — subtree first (to understand what
the research is actually about), then the shape diagnostics (to
understand the structural and distributional health), then the recent
calls in reverse-chronological order (most recent first). Produce a
structured punch list of problems worth fixing.
The shape diagnostics section contains auto-detected findings from
three scans: graph health (structural topology problems), rating shape
(credence/robustness distribution issues), and review signals
(aggregated self-reports from calls). Each finding has a severity and
suggested action. Incorporate relevant findings into your punch list —
don't just repeat them, but use them as evidence alongside your own
reading of the subtree and calls. You may disagree with or downgrade
a finding if the full context warrants it.
Format each punch list item as:
[severity 1-5] <short_id> <what's wrong in one sentence>
→ suggested: <action>
Where <action> is one of:
dispatch <call_type> — a rumil call should be run to address this
(e.g. dispatch assess to replace a stale judgement)
apply_move <MOVE_TYPE> — a cc-mediated move would fix this directly
(e.g. apply_move CREATE_QUESTION to add a missing sub-question, or
apply_move FLAG_FUNNINESS to flag a page for human review)
edit_prompt <filename> — the prompt the producing call used looks
like the root cause; it should be edited. Identify the specific file
in prompts/.
inspect — a human should read the full trace; the problem is real
but not clearly actionable
ignore — borderline; flagged for completeness but not worth
chasing
After the punch list, add a short shape-level summary (2-4 bullets):
patterns that cut across multiple items. Examples: "most find_considerations
calls on this question are producing thin output — suggests the task
prompt needs strengthening", or "the assess judgement is stale relative
to three considerations created after it".
Guidance
- Be specific. Every item must cite a concrete page or call by
short ID. Don't produce general observations without attribution.
- Always summarize what an ID refers to. When you cite
be6d1a1d,
write be6d1a1d (the AI-governance-determines-space-allocation claim),
not just the bare hex. Short IDs mean nothing without a quick gloss;
forcing the user to look them up slows every discussion. Applies to
page IDs, call IDs, and any other workspace entity.
- Rank ruthlessly. If you find 15 issues, put the three that
actually matter at the top. Severity 5 should be rare — reserve it
for problems that invalidate research output.
- Favor small fixes. A single
apply_move is cheaper than a full
re-dispatch, which is cheaper than editing a prompt and re-running
everything. Suggest the smallest fix that clears the issue.
- Don't dispatch or apply anything yourself. This skill is
read-only. The user picks what to act on from your punch list and
fires the relevant skills manually.
- If the confusion-scan verdicts are present, treat them as prior
evidence but don't blindly defer to them — you have the subtree
context they didn't. Disagree when warranted and say why.
Follow-ups the user may ask for
- "trace the top one" → suggest
/rumil-trace <call_id>
- "let's fix it" → suggest
/rumil-clean <question_id> for a guided
accreting-only cleanup, or individual /rumil-dispatch calls for
re-runs. For prompt-level fixes, pull the trace with /rumil-trace <call_id> and edit the relevant prompts/*.md directly.