best-practices

one kernel for shipping changes. one engineering layer underneath, one agent layer on top. small enough to reread, dense enough to matter. no fluff, no theater, no agreement for the sake of agreement.

works as a claude code skill and slash command, a codex skill and AGENTS.md, and a portable markdown kernel for cursor, antigravity, gemini cli, continue, cline, aider, custom GPTs, and raw API harnesses.

read first. write second. verify third.

---

the layers

each layer assumes the one below. flatten them and the hierarchy collapses.

the three layers are the meditation. the loop is the meditation made runnable: the same kernel as a loop you walk on one real goal, intent to undo.

---

engineering kernel

six cuts. three acts. these are the rules that make any of the rest possible. shipping-rules sits on top.

before

during

after

these six are the kernel. internalize them and the rest of the file reads as the consequences, not the rules.

---

agent kernel

shipping with help (yourself, a teammate, an agent, a swarm of agents) does not exempt rigor. it nests rigor inside coordination.

the full text lives in shipping-rules.md. the kernel is:

• one chair. every change has one human who owns the call.
• bounded slices. no overlapping write scopes, no implicit shared work.
• explorers map, workers implement, verifiers gate. roles are not labels, they are different read/write contracts.
• acceptance criteria written before execution. if you cannot write the bar, the slice is not ready.
• per-change rigor inside every slice. orchestration does not buy you out of the engineering kernel. it amplifies it.
• closeout has five parts. integrated result, verification summary, commit ids per slice, notes current, next slice with rationale. fewer means open.

agents produce plausible code that quietly does the wrong thing. humans do too. same rigor, no exceptions.

agents have one extra constraint humans do not: context is a budget, not a backdrop. degrade gracefully when full. clear when poisoned by failed approaches. dispatch fresh-context reviewers, not the same head twice. the reviewer who never wrote the code spots more than the writer who just finished it.

codex subagents

codex makes the agent kernel concrete with subagents. the main thread stays clean: requirements, decisions, integration, closeout. subagents take bounded noisy work: exploration, tests, triage, log analysis, or a slice with a disjoint write scope.

do not fan out by default. codex only spawns subagents when asked explicitly, and each subagent spends its own model and tool budget. write the acceptance bar before the spawn. ask for summaries, not raw tool spew.

---

the loop

the kernel is a meditation. the loop is the kernel made runnable: a fixed ten-checkpoint loop you point at one goal and walk, intent to undo, until the goal is met. it does not replace the cuts. it routes you through them.

read sits at checkpoint 01, write at 03, verify at 04, the undo plan at 09. the loop closes on itself: an undo plan is not optional. a central core holds the command center; each pass feeds fixed checkpoints instead of spawning new notes, so the map stays a clean circle no matter how much you gather.

the same rigor applies inside every checkpoint. acceptance criteria written before the work. evidence, with a source, before a claim enters the verify ledger. a bounded stop instead of an infinite loop. prune archives, it does not delete. the agent kernel applies when a checkpoint fans out: read dispatches explorers, verify dispatches fresh-context skeptics, the core is the one chair.

it ships as an obsidian vault you can see and navigate, plus a claude code skill that scaffolds and drives it. enforcement is still superpowers; the loop is operationalization. compose, do not substitute.

the live graph: ten checkpoints orbiting a single core, no orphans.

→ run it with /best-practices "" · loop/ · loop/references/orchestration.md

---

the stance

context over text. calibrated confidence. evidence over vibes. no agreement theater. confidence is earned, not asserted. skepticism is not new information. accountability is non-transferable: you read because you sign.

the stance is what makes the rest of this file work. without it, the kernel becomes a checklist, and a checklist is ceremony, not rigor. the rules are load-bearing only when the person reading them is already willing to be wrong in public.

---

install

clone first, copy what you need. nothing here uses curl, so it works against private and public repos identically.

git clone https://github.com/AgriciDaniel/best-practices.git
cd best-practices

then pick one or more paths.

codex: skill (auto-loads on relevant work)

mkdir -p ~/.codex/skills/best-practices/agents
cp SKILL.md ~/.codex/skills/best-practices/
cp agents/openai.yaml ~/.codex/skills/best-practices/agents/

codex reads ~/.codex/skills/best-practices/SKILL.md as the loadable skill. agents/openai.yaml is optional metadata for the skill list and default prompt. the skill is global: it gives codex the kernel when the work calls for it, but it does not make every repo inherit the kernel.

codex: project AGENTS.md (repo-local contract)

cp AGENTS.md /path/to/your/project/AGENTS.md

codex reads AGENTS.md at the repo root. use this when the project itself should carry the kernel for every codex session in that repository.

codex: hybrid install (recommended)

mkdir -p ~/.codex/skills/best-practices/agents
cp SKILL.md ~/.codex/skills/best-practices/
cp agents/openai.yaml ~/.codex/skills/best-practices/agents/
cp AGENTS.md /path/to/your/project/AGENTS.md

the hybrid shape is intentional: SKILL.md is the reusable global meditation, AGENTS.md is the repo-local operating contract. use both when you want codex to remember the kernel generally and carry it specifically inside a project.

for bigger work, add codex subagents on top: ask explicitly for parallel agents, keep read-heavy work off the main thread, and use separate write scopes or worktrees before letting multiple agents edit at once. codex surfaces active agent threads through /agent; use /fork only when the work truly branches.

claude code: skill (auto-loads on relevant work)

mkdir -p ~/.claude/skills/best-practices
cp SKILL.md ~/.claude/skills/best-practices/

claude code reads ~/.claude/skills/best-practices/SKILL.md and auto-injects the kernel when the description matches your prompt. copy SKILL.md only. do not cp -r the whole repo into the skill dir, that would drop .git and other non-skill files alongside SKILL.md. only SKILL.md belongs in the skill dir.

claude code: slash command (explicit invocation)

mkdir -p ~/.claude/commands
cp best-practices.md ~/.claude/commands/

then /best-practices injects the full kernel, and /best-practices injects just one section. ten sections addressable: stance, engineering, agent, loop, read, name, small, delete, evidence, failure. unrecognized arguments fall through to the full kernel with a single-line note.

claude code: the loop (run the kernel on a goal)

the loop ships inside this skill. for the kernel text alone, copying SKILL.md (above) is enough; to run the loop the loop/ scripts must be present, so symlink the clone as the skill:

ln -s "$(pwd)" ~/.claude/skills/best-practices

then /best-practices "" scaffolds and drives a ten-checkpoint loop for that goal as an obsidian vault (a section name still scopes the kernel; bare /best-practices prints the full kernel). see loop/references/orchestration.md.

claude code: project CLAUDE.md (shared with the team)

claude code reads CLAUDE.md at the repo root, not AGENTS.md. for a project-scoped contract, copy the kernel into CLAUDE.md:

cp AGENTS.md /path/to/your/project/CLAUDE.md

antigravity / gemini cli / openai agents / other AGENTS.md readers

these tools read AGENTS.md at the repo root by convention.

cp AGENTS.md /path/to/your/project/AGENTS.md

for antigravity, prefer AGENTS.md when your version supports it. if your setup still expects GEMINI.md or an antigravity-specific rules file, use a thin wrapper there that points at AGENTS.md, or paste the same kernel text.

cursor / antigravity / continue / cline / aider

each tool has its own rules file convention and the conventions move. drop the AGENTS.md content into your tool’s current path. starting points (verify against current docs before committing):

• cursor: .cursor/rules/best-practices.mdc (with frontmatter wrapper).
• antigravity: AGENTS.md at repo root when supported; otherwise a thin GEMINI.md or antigravity rules wrapper that points to AGENTS.md.
• continue: ~/.continue/config.json rules array, or older .continuerules file.
• cline: .clinerules (file or directory) at repo root.
• aider: pass with --read AGENTS.md or via .aider.conf.yml.

if your tool is not listed, AGENTS.md is plain markdown and works as a generic rules file in any harness that loads markdown.

---

how it composes

• needs enforcement for adversarial agents (rationalization guards, iron laws, red-flag stop-words) -> stack obra/superpowers
• needs iron-law TDD -> superpowers:test-driven-development
• needs debugging discipline -> superpowers:systematic-debugging
• needs parallel-agent SOP -> superpowers:dispatching-parallel-agents

this kernel is the meditation. those are the enforcement. compose, do not substitute.

---

what this is

a meditation. six axioms compressed into something you reread on a monday morning when you forgot why you ship the way you ship. it works for a person or an agent who already wants to be rigorous.

what this is not

• not a checklist. checklists rot. kernels compress.
• not a textbook. textbooks are for things you forget. this is for things you use.
• not exhaustive. exhaustive lists are the enemy of rereading. six cuts is the point.
• not an enforcement layer. there is no iron-law framing here, no rationalization-blocking tables, no red-flag stop-words. the kernel will not defend itself against an agent that decides to skip steps. for that, compose with obra/superpowers or another enforcement-grade ruleset.
• not a substitute for TDD discipline. cuts 5 and 6 imply tests, they do not mandate red-green-refactor. if you need iron-law TDD, install the superpowers test-driven-development skill on top of this kernel.
• not original in the sense of inventing axioms. original in the sense of picking the load-bearing few and naming them clearly.

---

license

MIT. fork it, rewrite it, ship it under your name. attribution appreciated, not required.

---

built by @AgriciDaniel · read first. write second. verify third.