godfly-skills

19 skills for Claude Code and Codex CLI that form one system: an adversarial reviewer, the evidence-grounded toolkit it draws on, the production-ops discipline it feeds into, and the gauntlet that proves any of it actually works.

Not a grab-bag. Every skill here has a deliberately carved territory and knows exactly which sibling to hand off to when the question isn't its job — and the whole analysis spine links to one shared evidence standard instead of restating it.

Why yet another skills repo? 🙃

Fair question. There are collections out there advertising hundreds of skills, toolkits with more of everything, and several published "adversarial reviewer" personas. Here is the honest pitch:

Most published skills are personas. These are protocols. A persona tells the model to "act like a tough reviewer." A protocol tells it what evidence it must gather before it is allowed to have an opinion, in what order, when it must yield, and what a finished verdict looks like. godfly's Steelman Guarantee is non-skippable: before challenging anything it restates your actual ask, your position at its strongest, and tests your broadest invariant before chasing narrower theories. A challenge against a position you don't hold is just noise with confidence.

No two skills fight over the same trigger. The most common failure of skill collections is five review skills that all activate on "review this." Territories here are deliberately carved:

Situation Skill
"Challenge this / poke holes / call my bullshit" godfly
Something is failing right now troubleshooting-investigator
Why did that failure really happen (post-hoc) root-cause
It works, but nobody understands it deep-dive
How could this fail in the future failure-analysis
Argue the strongest case for the other side devils-advocate
What are we taking for granted assumptions-check
Choose between competing approaches competing-hypotheses
Attack paths + launch readiness red-blue-review
Imagine the failure before starting / write it up after premortem-postmortem

Evidence is a shared standard, not a vibe. evidence-grounding is the canonical evidence-standards reference; the whole analysis spine links to it instead of restating it. Challenges carry evidence tiers, and the house rule is blunt: agreement without evidence is a bug.

The ops suite was forged in production, not written for a repo. These rules exist because their absence hurt:

  • mean-qa's Environment Boundary: an environment you cannot classify is production. A flow passes only when UI, network, and backend layers all confirm — UI-green alone is unproven, not pass.
  • incident-validator distinguishes "PR merged" from "verified running in production" and asks which one you actually proved. It grades restraint — probes you deliberately did NOT run against partner systems — as a quality gate.
  • safe-ops classifies every operation L1–L4 with dry-run previews, named confirmation gates, and audit trails.
  • toolshed keeps durable working state for one task and then deletes it at close — git history is the archive, and decisions promote to ADRs the moment they're made, not when someone remembers.

One library, two runtimes. The same skill files run in Claude Code (~/.claude/skills/) and OpenAI Codex CLI (~/.codex/skills/). Cross-references are relative, so the system stays intact wherever you drop it.

The system

flowchart TD
    Q{What's the situation?}
    Q -->|"challenge / review / stress-test"| G[godfly]
    Q -->|failing right now| TI[troubleshooting-investigator]
    Q -->|why did it happen| RC[root-cause]
    Q -->|works but opaque| DD[deep-dive]
    Q -->|risky operation| SO[safe-ops]
    Q -->|QA campaign| MQ[mean-qa]
    Q -->|watching a deploy| DM[deployment-monitor]
    Q -->|close an incident?| IV[incident-validator]
    Q -->|multi-session task state| TS[toolshed]

    G --> AC[assumptions-check]
    G --> CH[competing-hypotheses]
    G --> EG[evidence-grounding]
    G --> FA[failure-analysis]
    G --> RB[red-blue-review]
    G --> PP[premortem-postmortem]
    G --> DA[devils-advocate]

    RC --> TI
    RC --> FA
    PP --> RC
    IV --> G
    IV --> RC
    TS --> SO
    TS --> HO[handoff]
    TS --> SA[spec-adr-builder]
    TS --> DM
    DM --> SO
    EG -.evidence standard.- AC & CH & FA & DD

The skills

Flagship

Skill What it does
godfly Adversarial collaborator with a non-skippable Steelman Guarantee, evidence-tiered challenges, a live-state PR/incident review gate, a hold/yield rule where evidence decides — not authority — and a permanent verdict graph under docs/verdicts/ so a settled claim is never re-derived.

The adversarial & analysis spine

Skill What it does
evidence-grounding The canonical evidence standard: tiers, precedent search, claim validation. Everything else links here.
assumptions-check CIA Key Assumptions Check — surface and rate what's being taken for granted.
competing-hypotheses Analysis of Competing Hypotheses — map the solution space, let evidence eliminate.
failure-analysis FMEA, dependency chains, and real-world failure pattern matching.
devils-advocate The strongest possible case for the side nobody is arguing.
red-blue-review Attack paths vs. defense controls with a ship/block/spike gate.
premortem-postmortem Kill the project on paper before it starts; write the honest document after it dies for real.
root-cause Evidence-backed Five Whys from symptom to systemic cause.
troubleshooting-investigator Structured investigation for things failing now, with a ranked hypothesis table.
deep-dive Map a system that works but isn't understood; produce an evidence-backed brief.

The production-ops suite

Skill What it does
mean-qa Adversarial QA campaigns that find what a happy-path pass misses — oracle discipline, a hard prod-safety boundary, and a run directory of embedded screen evidence and replayable cases.
incident-validator Gate matrix for incident handovers, fix PRs, and postmortems — with live evidence verification and a closure verdict.
safe-ops L1–L4 risk classification, dry-run previews, named confirmation gates, audit trails.
deployment-monitor Read-only post-deploy evidence gathering, anomaly detection, and cadence summaries.
toolshed Mortal working state for one task under docs/work/<slug>/ — deleted at close, survivors become ADRs/specs.
handoff Compact continuation notes so the next session/agent doesn't reconstruct the thread.
spec-adr-builder Specs, ADRs, and RFCs with non-goals, alternatives, rollout, and rollback.
gauntlet Empirically tests whether a skill beats the raw model: planted-flaw fixtures, no-skill control arms, blind judging, pressure tests, cross-model runs via installed CLIs. Every finding becomes a permanent rule in the skill under test.

Install

Claude Code:

git clone https://github.com/CassioRoos/godfly-skills.git
mkdir -p ~/.claude/skills
cp -R godfly-skills/skills/* ~/.claude/skills/

Codex CLI:

mkdir -p ~/.codex/skills
cp -R godfly-skills/skills/* ~/.codex/skills/

Or symlink individual skills if you only want part of the system — but note the cross-references: godfly leans on the spine, toolshed leans on the ops suite.

Design principles

  1. Verdict first. Blocker, impact, evidence, action — never buried under preamble.
  2. Evidence decides, not authority. Hold when the evidence holds; yield cleanly when it doesn't.
  3. Steelman before challenge. Prove you understood the position before attacking it.
  4. Unclassifiable environment = production. Ceremony scales with blast radius.
  5. State is mortal, decisions are not. Working notes die at task close; contract-changing decisions promote to ADRs the moment they're made; settled cross-task claims persist as verdict nodes in docs/verdicts/.
  6. Aim at the code, never the person. Heat scales with stakes, and it's always pointed at the work.

License

MIT