How I use AI through a repeatable and programmable workflow to stop fixing the same mistakes over and over

Quick context: I use AI heavily in daily development, and I got tired of the same loop.

Good prompt asking for a feature -> okay-ish answer -> more prompts to patch it -> standards break again -> rework.

The issue was not "I need a smarter model." The issue was "I need a repeatable process."

The real problem

Same pain points every time:

• AI lost context between sessions
• it broke project standards on basic things (naming, architecture, style)
• planning and execution were mixed together
• docs were always treated as "later"

End result: more rework, more manual review, less predictability.

What I changed in practice

I stopped relying on one giant prompt and split work into clear phases:

1. /pwf-brainstorm to define scope, architecture, and decisions
2. /pwf-plan to turn that into executable phases/tasks
3. optional quality gates:
   • /pwf-checklist
   • /pwf-clarify
   • /pwf-analyze
4. /pwf-work-plan to execute phase by phase
5. /pwf-review for deeper review
6. /pwf-commit-changes to close with structured commits

If the task is small, I use /pwf-work, but I still keep review and docs discipline.

The rule that changed everything

/pwf-work and /pwf-work-plan read docs before implementation and update docs after implementation.

Without this, AI works half blind. With this, AI works with project memory.

This single rule improved quality the most.

References I studied (without copy-pasting)

• Compound Engineering
• Superpowers
• Spec Kit
• Spec-Driven Development

I did not clone someone else's framework. I extracted principles, adapted them to my context, and refined them with real usage.

Real results

For me, the impact was direct:

• fewer repeated mistakes
• less rework
• better consistency across sessions
• more output with fewer dumb errors

I had days closing 25 tasks (small, medium, and large) because I stopped falling into the same error loop.

Project structure that helped a lot

I also added a recommended structure in the wiki to improve AI context:

• one folder for code repos
• one folder for workspace assets (docs, controls, configs)

Then I open both as multi-root in the editor (VS Code or Cursor), almost like a monorepo experience. This helps AI see the full system without turning things into chaos.

Links

Repository: https://github.com/J-Pster/Psters_AI_Workflow

Wiki (deep dive): https://github.com/J-Pster/Psters_AI_Workflow/wiki

If you want to criticize, keep it technical. If you want to improve it, send a PR.