{"slug": "show-hn-power-coding-an-ai-coding-skill-i-audited-on-claude-code-and-codex", "title": "Show HN: Power Coding – an AI-coding skill I audited on Claude Code and Codex", "summary": "Anand Pareek introduces Power Coding, a structured framework for building scalable apps with AI coding agents that installs five counter-disciplines—borrowed from aerospace, the Toyota Production System, and eval-driven ML practice—as plain markdown files readable by any agent (Claude Code, Codex, Gemini CLI). The framework addresses five forces that kill AI-assisted projects: ambiguity, amnesia, drift, opacity, and entropy, with counter-forces including alignment, memory, the loop, sight, and foresight.", "body_md": "*A structured framework for building scalable apps with AI coding agents, by Anand Pareek.*\n\nFive forces quietly kill every AI-assisted project. Each one has a proven counter-discipline — borrowed from aerospace, the Toyota Production System, and modern eval-driven ML practice, and translated for LLM agents. Power Coding installs all five as one system, in plain markdown files your agent re-reads every session. Works with any agent (Claude Code, Codex, Gemini CLI…) — no daemons, no hooks, no lock-in.\n\n| # | Force (what kills projects) | Counter-force | What gets installed |\n|---|---|---|---|\n| 1 | Ambiguity — the agent confidently builds the wrong thing |\nAlignment |\nA one-page brief (`PRD.md` ) every eval and scan reads · plain-language architecture sign-off before code is built against it (and before every shape-changing enhancement) · a decision log that's never silently reversed · a scope brake (smallest version that proves it works, first) |\n| 2 | Amnesia — every session starts from zero |\nMemory |\n`Handoff.MD` — one magic phrase chains sessions (\"Refer to Handoff.MD in ), auto-updated before context runs out · `<path>` and begin\"`Learning.MD` — every bug gets a 5-whys root-cause entry so no mistake is made twice · git checkpoints at every working state and before every risky change — rollback beats a debugging spiral |\n| 3 | Drift — \"done\" is declared without proof |\nThe Loop |\n`Loop.MD` free evals drafted from your brief, run for real after every change with per-eval ✅/❌ · a golden set for agentic apps (curated inputs + reference outputs + check method, drafted from the brief) that runs only with your approval — paid API calls are never spent automatically · a status machine that offers to turn the loop on the moment the first draft works · evals that pass 3 runs in a row graduate into real automated tests |\n| 4 | Opacity — you can't see what was built, or where the effort went |\nSight |\nMermaid flow diagrams to a strict debuggable convention (one step per box, AGENT vs FUNCTION, gates with real thresholds) + a standalone HTML viewer + optional local-only in-app debug tab — debug by pointing at a box instead of reading code · Session Pulse: a 2-line effort split (🎯 feature · 🔧 support · 🔁 rework) after major milestones |\n| 5 | Entropy — silent failures accumulate until production finds them |\nForesight |\nFMEA scans cross-referencing your diff against the PRD and architecture (Severity×Occurrence×Detection → RPN → P0/P1/P2, mandatory 12-category walkthrough) · Sentinel Scan: a silent four-lens sweep after every major completion (misuse traps, stop signals, silent failures, waste) — one-line flags only when something fires · a day-one security baseline (no committed secrets, server-side enforcement, validated inputs) |\n\nNone of this is invented from scratch — each counter-force is a reliability discipline with decades of production behind it, applied to LLM coding:\n\n**FMEA**— aerospace/automotive failure-mode analysis** 5-whys, Poka-Yoke, Andon, Jidoka, Muda**— the Toyota Production System (they power Learning.MD and Sentinel's four lenses)** OEE**from Total Productive Maintenance — the insight behind Session Pulse: your bottleneck is almost never what you think it is**The Loop**— Andrew Ng's agentic-workflow argument: output quality comes from the build→check→revise loop more than from the model\n\nEvery behavior fits one of three tiers — so the framework helps without nagging:\n\n**Cheap, append-only**(Sentinel, Pulse, file updates) → you opt in once at setup, then it runs silently. No per-run popups for cheap things — question fatigue trains yes-clicking, which destroys the questions that matter.**Expensive**(full FMEA scan, golden-set / paid eval runs) → asks before running; configurable to auto. Turning the loop on consents only to the free evals — paid runs are always consented separately.**State-mutating**(git commits, architecture changes) → commits are auto-with-announcement or propose-first (your pick at setup); architecture changes**always** wait for your approval — no config can silence that.\n\nAgents only reliably re-read two things: their platform instructions file and the files it points to. Setup writes one pointer block into `CLAUDE.md`\n\n/ `AGENTS.md`\n\n/ `GEMINI.md`\n\n, and every \"automatic\" behavior is a written instruction in `Handoff.MD`\n\n/ `Learning.MD`\n\n/ `Loop.MD`\n\nthat the pointer forces back into context — with deliberate redundancy so triggers survive across sessions and across models. No scheduler exists; the agent is the runtime.\n\n**Claude Code:**\n\n```\ngit clone https://github.com/APareek89/power-coding.git ~/.claude/skills/power-coding\n```\n\n**Other agents (Codex, Gemini CLI, …):** clone anywhere and point your agent's skill/instructions mechanism at `SKILL.md`\n\n.\n\n`power-coding`\n\n— first invoke in a repo runs setup (a few short questions, asked once); later invokes resume ongoing duties`power-coding brief`\n\n— create/update the one-page PRD (if you skipped it at setup)`power-coding fmea`\n\n— run an FMEA scan now ·`fmea-config`\n\n— change when it runs`power-coding sentinel`\n\n— run a Sentinel Scan now ·`sentinel-config`\n\n— change when it runs`power-coding pulse`\n\n— show the Session Pulse now ·`pulse-config`\n\n— change when it shows\n\n```\nSKILL.md                    the orchestrator (start here)\nreferences/                 full protocols: loop-engineering, fmea, diagram convention, debug-tab recipes\nassets/                     templates: PRD.md, Handoff.MD, Learning.MD, Loop.MD, module config + state\nscripts/ (node, dep-free)   secret-scan (pre-commit gate) · sentinel (change-set/staleness/drift detection) · mermaid validate + HTML viewer\n```\n\n", "url": "https://wpnews.pro/news/show-hn-power-coding-an-ai-coding-skill-i-audited-on-claude-code-and-codex", "canonical_source": "https://github.com/APareek89/power-coding", "published_at": "2026-07-24 14:11:45+00:00", "updated_at": "2026-07-24 14:22:30.103677+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-research", "ai-products"], "entities": ["Anand Pareek", "Claude Code", "Codex", "Gemini CLI", "Toyota Production System", "Andrew Ng"], "alternates": {"html": "https://wpnews.pro/news/show-hn-power-coding-an-ai-coding-skill-i-audited-on-claude-code-and-codex", "markdown": "https://wpnews.pro/news/show-hn-power-coding-an-ai-coding-skill-i-audited-on-claude-code-and-codex.md", "text": "https://wpnews.pro/news/show-hn-power-coding-an-ai-coding-skill-i-audited-on-claude-code-and-codex.txt", "jsonld": "https://wpnews.pro/news/show-hn-power-coding-an-ai-coding-skill-i-audited-on-claude-code-and-codex.jsonld"}}