slopflow

An opinionated, evidence-first discipline layer for Claude Code and Codex: the antidote to AI slop.

Coding agents are fast, fluent, and confident. That confidence is the problem: they infer behavior from filenames, trust stale comments, claim things are "already implemented," collapse inference into fact, and optimize for sounding decisive over being right. slopflow installs a shared set of always-on rules plus a skill library that force the opposite habits: ground claims in executable evidence, trace the real production path, reason before converging, and preserve uncertainty instead of hiding it.

What it installs

slopflow ships two layers:

1. Always-on prompt: global instructions loaded into every session.

   • Claude → ~/.claude/CLAUDE.md
   • Codex → ~/.codex/AGENTS.md
   • These hold the Core Rules (evidence before claims, production-path first, fact vs inference, signal-loss and metric/oracle audits, meta-level reasoning checks), skill routing, identities, and the default response shape.

2. Skills: focused workflows loaded on demand when their triggers match. Installed to ~/.claude/skills/<skill>/ and ~/.codex/skills/<skill>/. Detailed procedure lives here so the always-on prompt stays lean.

Install

git clone https://github.com/1ikeadragon/slopflow.git
cd slopflow
./install.sh

The installer must be run from a complete checkout (it checks for CLAUDE.md, AGENTS.md, and skills/).

Flag	Effect
--dry-run	Print every action without touching the filesystem.
--claude-only	Install only the Claude target.
--codex-only	Install only the Codex target.
-h, --help	Usage.

Environment overrides:

Variable	Default	Purpose
CLAUDE_HOME	~/.claude	Claude install root.
CODEX_HOME	~/.codex	Codex install root.
SLOPFLOW_BACKUP_ROOT	~/.slopflow-backups/<timestamp>	Where overwritten files are backed up.

Safety and backups

The installer never silently destroys your existing config:

• Any prompt file or same-named skill it is about to overwrite is first copied to ~/.slopflow-backups/<timestamp>/, mirroring its original path.
• Stale in-place skill backups named *.bak.* are moved out of the skills root into the backup directory, so they are not loaded as skills.
• The retired security-review skill is removed (after being backed up).

Preview everything first with ./install.sh --dry-run.

Uninstall / restore

There is no uninstaller. To revert, restore the relevant files from the most recent ~/.slopflow-backups/<timestamp>/ directory (it mirrors the original paths), and delete any slopflow skill directories you no longer want from ~/.claude/skills/ and ~/.codex/skills/.

Skills

Skill	Use it for
reasoning-discipline	The reasoning spine, applied by default to non-trivial / ambiguous / high-stakes work: frame → generate options → verify → diagnose → evaluate → decide → synthesize, plus meta-level self-checks.
evidence-first	Grounding any non-trivial claim in executable evidence; classifying code as production / test / legacy; preserving uncertainty.
implementation-discipline	Making the smallest correct change, with a plan + plan-audit before coding and a post-implementation audit after.
architecture-review	Mapping active wiring, entrypoints, data/control flow, runtime variants, and failure modes before proposing a design.
code-review-discipline	Hunting regressions, hidden callers, edge cases, error gaps, and weak tests. Correctness over politeness.
adversarial-test-design	Invariant-driven tests that attack assumptions, for security-sensitive, correctness-critical, parser, workflow, and AI/LLM logic.
legacy-cleanup	Classifying code by reachability before deletion; quarantine/delete decisions with validation.
rca-investigation	Cumulative, proof-driven root-cause analysis with competing hypotheses and code/config/DB/git tracing.
workflow-rca	Collecting workflow/runtime evidence (Temporal/CI IDs, cloud logs, artifacts, cache state) to feed rca-investigation.
rigorous-web-research	Current external research with hard citations: competitor practice, primary docs, recent papers, benchmarks.

The installer auto-discovers every directory under skills/ that contains a SKILL.md, so adding a skill is just adding a folder.

Philosophy

• Evidence before claims. Filenames, comments, and READMEs are hints, not proof.
• Production path first. Trace from a real entrypoint before judging or changing behavior.
• Separate fact from inference. Observed / Inferred / Uncertain / Conclusion.
• No hidden signal loss. Call out every shortcut, mock, fallback, or truncation and its risk.
• Reason before converging. Frame the problem, generate real alternatives, decide deliberately.
• Preserve uncertainty. A correct uncertain answer beats a confident false one.

License

MIT