jcosta33/skills

revolver

Run exhaustive multi-angle review and sequential repair across code, diffs, artifacts, plans, or systems. Use when breadth across every target-justified stance is required. Do not use for fast fixed-panel review or single-claim verification.

First seen Jul 12, 2026

Installation

$ npx skills add jcosta33/skills --skill revolver

Similar popular skills

Related neighbors and high-traction skills in the same topics — useful to compare before installing.

Also in this package

Other skills from jcosta33/skills · top by installs.

npx skills add jcosta33/skills

Browse all from jcosta33/skills

More details

Agent compatibility

Declared targets from SKILL.md / docs. Unmarked agents are not listed — the skill may still install via the CLI.

Claude Code Not declared
Cursor Not declared
Codex Not declared
GitHub Copilot Not declared
Windsurf Not declared
Gemini CLI Not declared
Cline Not declared
OpenCode Not declared

Repository health

Stars 3
License LICENSE
Default branch main
Open issues 0
Status Active

Package contents

Files included with this skill beyond the listing page.

  • skill md SKILL.md 2,435 B
  • docs SUMMARY.md 257 B

History

  1. First seen on skills.sh
  2. First recorded snapshot · 39 installs

SKILL.md

Revolver

Every justified angle. Every finding dead before the next.

Method

  1. Freeze the target state.
  2. Derive the complete materially distinct stance pool from requirements, trust boundaries, data

and control flow, changed surfaces, failure modes, users, and operating conditions. Cover every exposed material risk. Reject duplicates and filler. Honor explicit human constraints; otherwise choose the pool without asking. Use no fixed floor or ceiling.

  1. Give each stance one falsification question and target evidence. Order the pool by consequence

and uncertainty. Freeze it for the rotation.

  1. Dispatch one fresh reviewer at a time on the cheapest adequate model. Give it the current target,

one stance, and no prior reviewer prose. If fresh dispatch is unavailable, stop. Self-review is not fresh.

  1. Fact-check every finding against direct evidence. Kill every finding before the next stance:

- supported: apply the fix and verify it when settled authority determines the fix; otherwise stop as a human decision; - refuted: reject it with decisive evidence; - human decision: force material ambiguity into explicit human selection; stop until selected; - unverified or blocked: name the missing evidence or capability and stop.

  1. Freeze the addressed target for the next stance.
  2. Finish the full pool. After a productive rotation, rebuild the pool against the current target.

Add or remove stances only between rotations.

  1. Stop after one quiet rotation or three cycles.

One cycle is one full pool rotation. Productive means a supported finding was fixed and verified. Quiet means no actionable finding survived verification.

Output

Return material fixes and verification, consequential refutations, and unresolved human decisions. For each decisive run, include command, working directory, state identifier, numeric exit, and the smallest untouched decisive output excerpt once.

Boundaries

Reviewers attack claims, callers, and companion surfaces. They return only actionable findings: location, consequence, evidence, minimal fix. The orchestrator fixes; reviewers stay independent.