Trigger only on explicit "ultra review" or "ultra-review".
SkillScore68Maintained
Automatic
◆Coding pack
Last verified 8/20/2026
What it changes
Passed 12 of 17 test cases with the skill, and 5 of 17 without it.
Without skill32%
With skill68%
Checks it passed that the model alone missed
Assigns 1 reviewer to Author A.
Assigns at least 4 reviewers to Author B.
Formats the timestamp folder as 20251024-1430.
and 13 more in the full report
10 fixed5 passed either way5 still failing2 broke
Pass-rate lift
+36 pp
Tokens
+313%
Agent turns
0%
Response time
-59%
Model gemini-3.6-flashMeasured 8/20/2026v115 of 22 cases passed with skill · 17 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 10
Prompt
The orchestrator finished generating `SUMMARY.md` in interactive mode. Before creating any tasks in our project management tool or writing `TASKS.md`, what action must the orchestrator take?
What was checked, without and with the skill
Checked
Without skill
With skill
Presents SUMMARY.md findings in chat prior to creating external tasks.
Missed
Passed
Uses AskUserQuestion to obtain user confirmation before creating external tasks.
Missed
Passed
case-20 · gemini-3.6-flash · 8/20/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md461 linessingle file
Ultra Review Workflow
Orchestrate a multi-round, multi-agent deep review. Goal: depth over speed. Maximize real problems found, minimize false positives via adversarial cross-critique.
The orchestrator never writes findings. Design the review, spawn agents, digest results.
Invocation modes
Interactive (user typed "ultra review"): run STEP 1 interview. STEP 10 offers either [PRODUCT_MANAGEMENT_TOOL] tasks or TASKS.md.
Autonomous (team-lead invocation at feature completion): the caller supplies scope, risk hotspots, size, and confidence policy directly. Skip STEP 1's user questions. Skip STEP 10's chat presentation. Always write TASKS.md; never write to [PRODUCT_MANAGEMENT_TOOL]. Return the TASKS.md path to the caller.
Core principles
Depth, not token-saving. Spend whatever it takes. Use Explore subagents for code search, MCP tools (Stripe, [PRODUCT_MANAGEMENT_TOOL], Aspire, DB, etc.), and any other tooling. Prefer Perplexity over WebSearch for online research.
Generic by design. Roster, clusters, and focus areas are co-designed with the user every time. No fixed catalog.
80/20. Agents spend ~80% on the few highest-risk areas (deep dive), ~20% sanity-scanning the rest.
No cap on agent count. Match the change: a 20K-line PR touching money, integrations, and migrations may warrant 25+ agents; a small UI fix may warrant 4.
Multiple agents on hot areas. Assign 3-4 agents to each high-risk area from different angles. Overlap is expected and strengthens signal.
Agents run independent and parallel. No coordinating, no splitting work mid-flight.
False-positive hunt. Round 2 disproves findings, not validates them. Survivors must be defensible.
Confidence is categorical, not percentages:
Certain — verified; can reproduce, cite exact code paths, quote evidence.
Likely — strong evidence, one step short of full verification. Must state what would close the gap.
Possible — plausible from patterns or partial evidence; not verified. Must state the gap.
Confidence policy (set in interview):
Default policy: "Certain only" (drop weaker) or "Allow Likely and Possible with explanation" (keep weaker findings, each with a "Why not Certain" note).
High-impact exception: agents flagged high-impact (security, data loss, privacy, regulatory) may always keep Likely and Possible findings with an uncertainty note, regardless of default. Better to surface a possible breach than drop it.
Output structure
All artifacts live in .workspace/<branch-name>/ultra-review/<YYYYMMDD-HHmm>/:
.workspace/<branch-name>/ultra-review/<timestamp>/
├── CONTEXT.md # Diff summary, [feature]/[task] excerpts, env, scope
├── ROSTER.md # Final agent list with affinity clusters
├── round1/
│ └── <agent-slug>.md # One file per agent
├── round2/
│ ├── ASSIGNMENT.md # Reviewer-to-author table from Round 1 triage