Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
46 changes: 46 additions & 0 deletions documents/patterns/feedback-loop.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
---
authors: [nitsanavni]
related_patterns:
- playgrounds
- chain-of-small-steps
---

# Feedback Loop

## Pattern
Set up automated feedback, give AI permission to iterate autonomously until goal is reached.

1. Identify clear success signal (tests pass, UI matches design, coverage hits X%)
2. Give AI access to that signal (tests, devtools, linter, logs)
3. Grant explicit permission: "Keep iterating until tests pass"
4. Step away - AI checks its own work and keeps refining

Human elevates from tactical executor to strategic director.

## Example

**Bug fix with DevTools MCP:**
- Goal: Fix layout bug in web app
- Signal: DevTools screenshots showing the visual issue
- Setup: Point AI to devtools
- Permission: "Keep trying until the layout matches the design"
- AI iterates: change code → check screenshot → adjust → repeat

**Debugging with console logs:**
- Goal: Understand why feature behaves incorrectly
- Signal: Console output showing timing, state, assumptions
- Setup: "Add console.log to help yourself understand what's wrong"
- Permission: "Debug this until you find the root cause"
- AI iterates: add logs → run → validate assumptions → adjust → repeat

**Getting tests to pass:**
- Goal: All tests green
- Signal: Test runner output
- Permission: "Fix the failing tests, keep running them until they all pass"
- AI iterates: read failure → fix code → run tests → repeat

**CI monitoring:**
- Goal: Green CI build
- Signal: Direct access to CI state/logs (e.g., via `gh` CLI)
- Permission: "Monitor CI and fix issues until build passes"
- AI iterates: check status → read logs → fix → push → check again