Jidoka Instruction Architecture

Build quality into agent instructions

One instruction architecture for humans and agents. Eight layers each hold a single job. Checks stop the line the moment a layer drifts.

Scroll

Instructions sprawl. Nothing stops the drift.

Prompts pile up. Layers restate each other. Jobs go stale. Nobody notices until an agent misbehaves. Jidoka takes the Toyota path and builds quality into the process itself. In one layered architecture, every layer owns a single job and carries a machine-checkable budget. A defect then traces to exactly one layer. The line stops before the defect ships.

8
Layers
Most general to most specific
1
Job per layer
No layer restates another
0
Guesswork
Every defect localizes to one layer

Eight layers. Most general to most specific.

Each layer loads at the right moment and owns one concern. Auto-loaded layers stay budgeted, so context never bloats. On-demand layers disclose only when the work calls for them.

L0
System Prompt

Harness mechanics: turns, tool calls, the completion signal. Nothing about your project.

L1
CLAUDE.md

Project identity: what it is, who it serves, and where to find its jobs and checklists.

L2
CONTRIBUTING.md & JTBD.md

Contribution standards and the jobs each persona hires the work to do.

L3
Agent Profile

One persona: voice, skill routing, and scope constraints. It sets boundaries. It does not give steps.

L4
Agent References

Cross-cutting protocols shared across agents: memory, coordination, approval.

L5
Skill Procedure

The complete, imperative steps for one domain of work. They need no tribal knowledge.

L6
Skill References

The data a procedure consults: templates, worked examples, lookup tables.

L7
Checklists

Binary verification at a pause point. It confirms. It does not explain.

What agents align to, and how alignment holds.

Three well-publicized ideas answer the halves of the problem. Together they explain why the layers have the shape they have. They also explain why the checks stop the line.

Toyota
Jidoka

How quality holds. The process builds in quality. Inspection afterward does not. The checks halt at the first defect. They never pass one downstream.

Christensen & Moesta
Jobs To Be Done

What agents align to. Every layer traces to the progress a persona seeks in a specific circumstance. No layer traces to a feature list.

Atul Gawande
The Checklist Manifesto

How alignment holds under load. Structured instructions keep humans and agents consistent in how they apply existing knowledge.

Two gates. One at entry, one at exit.

Checklists never teach. They confirm. If you must explain an item, the procedure above it is incomplete. A semantic <read_do_checklist> or <do_confirm_checklist> tag wraps each gate. Every pause point in the repository is then one rg search away. You need no map.

Entry gate
READ-DO

Read each item, then do it. The gate loads constraints into memory before the first line of work. At that moment, one missed constraint sends everything in the wrong direction.

rg '<read_do_checklist'
Exit gate
DO-CONFIRM

Do from memory, then pause and confirm. The gate verifies that you missed nothing before a commit, merge, or release. The checks stay independent. They do not interrupt you mid-flow.

rg '<do_confirm_checklist'

Adopt it in three lines.

Install the skill pack. Tell Claude to set it up.

Terminal
cd my-repo/
apm install forwardimpact/jidoka-skills
echo "Set up Jidoka" | claude

Then wire the jidoka binary (or npx @forwardimpact/jidoka) into your checks, so the line stops at the first drifted layer. Read the full standard in the layered instruction architecture guide.