Feedback Loop
C
03 · Tech Stack (Rawclaw)

Feedback Loop

Planner → Generator → Evaluator.

Last synced 2026-04-28

Feedback Loop (read this before every task)

Your outputs are automatically grounded and graded.

Before you write: Retrieved context (corrections, voice snippets, copy examples) is injected into your system prompt by worker-cli before you see the task brief. You don't call retrieval yourself — it happens upstream.

After you write: An evaluator agent scores your output against:

  • Retrieved corrections relevant to this task
  • Voice corpus match (Chris-voice alignment)
  • Task success criteria from the brief

If it fails: You receive a feedback prefix listing only the specific failures. You revise. Max 3 loops, then it escalates to Chris via Telegram.

If Chris edits your output: His edit is captured in the corrections table via /correct and will be injected into your context on similar future tasks. You don't need to do anything — the system handles it.

What this means for you:

  • Write against the task brief. That brief includes success_criteria — match them.
  • Lean on retrieved corrections. They represent past fixes Chris made to similar work.
  • When you receive a feedback prefix, treat each bullet as a concrete revision instruction.
  • Never include your evaluator score or loop count in your output.