Power Coding — The Five Forces of Mindful Coding

A structured framework for building scalable apps with AI coding agents, by Anand Pareek.

Five 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.

Proof — measured, not claimed

Two independent audits of real production builds — one over Claude Code sessions, one over Codex sessions — are written up here. Honest summary:

On a large, failure-prone build, ~4–8 saved agent-hours and ~3–6M non-cached tokens per build.
The FMEA / Sentinel gates caught ≥32 real defects before they shipped — including a cross-tenant IDOR, SSRF / DNS-rebinding exposure, duplicate-queue & approval races, and non-fail-closed auth middleware.
It's an honest counterfactual estimate, not an A/B test — and on tiny, low-risk changes it's not worth the overhead. The audit keeps every limitation in.

The Five Forces

#	Force (what kills projects)	Counter-force	What gets installed
1	Ambiguity — the agent confidently builds the wrong thing	Alignment	A 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)
2	Amnesia — every session starts from zero	Memory	`Handoff.MD` — one magic phrase chains sessions ("Refer to Handoff.MD in `<path>` and begin"), auto-updated before context runs out · `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
3	Drift — "done" is declared without proof	The Loop	`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
4	Opacity — you can't see what was built, or where the effort went	Sight	Mermaid 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
5	Entropy — silent failures accumulate until production finds them	Foresight	FMEA 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)

The lineage

None of this is invented from scratch — each counter-force is a reliability discipline with decades of production behind it, applied to LLM coding:

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

The consent model

Every behavior fits one of three tiers — so the framework helps without nagging:

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.

How it works (the whole trick)

Agents only reliably re-read two things: their platform instructions file and the files it points to. Setup writes one pointer block into CLAUDE.md / AGENTS.md / GEMINI.md, and every "automatic" behavior is a written instruction in Handoff.MD / Learning.MD / Loop.MD that 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.

Install

Claude Code:

hljs language-bash

git clone https://github.com/APareek89/power-coding.git ~/.claude/skills/power-coding

Other agents (Codex, Gemini CLI, …): clone anywhere and point your agent's skill/instructions mechanism at SKILL.md.

Use

power-coding — first invoke in a repo runs setup (a few short questions, asked once); later invokes resume ongoing duties
power-coding brief — create/update the one-page PRD (if you skipped it at setup)
power-coding fmea — run an FMEA scan now · fmea-config — change when it runs
power-coding sentinel — run a Sentinel Scan now · sentinel-config — change when it runs
power-coding pulse — show the Session Pulse now · pulse-config — change when it shows

Layout

hljs language-sql

SKILL.md                    the orchestrator (start here)
references/                 full protocols: loop-engineering, fmea, diagram convention, debug-tab recipes
assets/                     templates: PRD.md, Handoff.MD, Learning.MD, Loop.MD, module config + state
scripts/ (node, dep-free)   secret-scan (pre-commit gate) · sentinel (change-set/staleness/drift detection) · mermaid validate + HTML viewer

Power Coding — The Five Forces of Mindful Coding

A structured framework for building scalable apps with AI coding agents, by Anand Pareek.

Proof — measured, not claimed

Two independent audits of real production builds — one over Claude Code sessions, one over Codex sessions — are written up here. Honest summary:

On a large, failure-prone build, ~4–8 saved agent-hours and ~3–6M non-cached tokens per build.
The FMEA / Sentinel gates caught ≥32 real defects before they shipped — including a cross-tenant IDOR, SSRF / DNS-rebinding exposure, duplicate-queue & approval races, and non-fail-closed auth middleware.
It's an honest counterfactual estimate, not an A/B test — and on tiny, low-risk changes it's not worth the overhead. The audit keeps every limitation in.

The Five Forces

#	Force (what kills projects)	Counter-force	What gets installed
1	Ambiguity — the agent confidently builds the wrong thing	Alignment	A 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)
2	Amnesia — every session starts from zero	Memory	`Handoff.MD` — one magic phrase chains sessions ("Refer to Handoff.MD in `<path>` and begin"), auto-updated before context runs out · `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
3	Drift — "done" is declared without proof	The Loop	`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
4	Opacity — you can't see what was built, or where the effort went	Sight	Mermaid 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
5	Entropy — silent failures accumulate until production finds them	Foresight	FMEA 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)

The lineage

None of this is invented from scratch — each counter-force is a reliability discipline with decades of production behind it, applied to LLM coding:

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

The consent model

Every behavior fits one of three tiers — so the framework helps without nagging:

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.

How it works (the whole trick)

Install

Claude Code:

hljs language-bash

git clone https://github.com/APareek89/power-coding.git ~/.claude/skills/power-coding

Other agents (Codex, Gemini CLI, …): clone anywhere and point your agent's skill/instructions mechanism at SKILL.md.

Use

power-coding — first invoke in a repo runs setup (a few short questions, asked once); later invokes resume ongoing duties
power-coding brief — create/update the one-page PRD (if you skipped it at setup)
power-coding fmea — run an FMEA scan now · fmea-config — change when it runs
power-coding sentinel — run a Sentinel Scan now · sentinel-config — change when it runs
power-coding pulse — show the Session Pulse now · pulse-config — change when it shows

Layout

hljs language-sql

SKILL.md                    the orchestrator (start here)
references/                 full protocols: loop-engineering, fmea, diagram convention, debug-tab recipes
assets/                     templates: PRD.md, Handoff.MD, Learning.MD, Loop.MD, module config + state
scripts/ (node, dep-free)   secret-scan (pre-commit gate) · sentinel (change-set/staleness/drift detection) · mermaid validate + HTML viewer

power-coding

Power Coding — The Five Forces of Mindful Coding

Proof — measured, not claimed

The Five Forces

The lineage

The consent model

How it works (the whole trick)

Install

Use

Layout

Similar Packages

power-coding

Power Coding — The Five Forces of Mindful Coding

Proof — measured, not claimed

The Five Forces

The lineage

The consent model

How it works (the whole trick)

Install

Use

Layout

Similar Packages