Go Loop

/go-loop — Iteration Loop Closer

Boris Cherny's /go pattern applied to AI media pipelines: when invoked, Claude tests its own work end-to-end, simplifies, commits, and pings you. No babysitting.

The Loop

GENERATE → VERIFY → [regen if FAIL, max 2 passes] → COMPOSITE → COMMIT → NOTIFY

When to Invoke

Invoke at the end of any generation task. Safe to chain after batch generators, direct API calls, or MCP tool calls.

Example prompts:

• regen shot-06 with updated prompt /go-loop
• generate remaining CGI shots using approved refs /go-loop
• try shot-08 variants for energy match /go-loop

Execution

Step 1 — Verify what was just generated

Run your verification script (see the verify skill for a CLIP-based implementation):

python3 scripts/verify.py <ITEM_ID> --json

Read the verdict JSON. If Claude just generated multiple items, iterate.

Step 2 — Regen if FAIL (max 2 passes)

For any item with verdict FAIL:

• Look at failed_metrics — targeted fix per metric:
  • character fail → ensure character refs are passed to the generator
  • palette fail → check world/palette assignment, enforce palette hint in prompt
  • prompt_match fail → revise prompt wording, tighten focal length
  • composite fail → cross-check frame selection
• Regen via same path. Do NOT burn credits if a prompt fix alone would resolve.
• Re-verify. If still FAIL after 2 regens → surface to user, do NOT continue loop.

Step 3 — Build composite

python3 scripts/assemble.py --composite <ITEM_ID>

Step 4 — Simplify (optional)

Run /simplify on any supporting code Claude wrote (prompt builders, batch scripts, helper utilities). Skip for pure generation work.

Step 5 — Commit

Summary commit with:

• Items touched (IDs)
• Pre/post quality scores
• Regen count per item
• References used

Commit message format:

<item-ids>: <short description>

Quality deltas: shot-06 0.31→0.42 (GOOD), shot-08 0.28→PASS, shot-11 FAIL after 2 regens

Co-Authored-By: Claude <noreply@anthropic.com>

Step 6 — Notify (optional)

Send summary to the user via your preferred channel (Telegram, Slack, email). Format:

/go-loop complete

<N> items verified, <M> regenned
Results:
  shot-06: GOOD (0.42)
  shot-08: PASS (0.31)
  shot-11: FAIL after 2 passes — needs manual review

Composite: runs/composites/latest.png
Commit: <hash>

Only send if the channel is configured. Otherwise skip silently.

Invariants (don't break these)

1. Never regen more than 2x per item per invocation. Burns credits; if still failing, the prompt or reference is the issue — hand back to user.
2. Never skip verification. The whole point is the tight loop. No verification = no /go.
3. Always pass references to the generator. Double-check before regen.
4. Respect file locks. If /lockscenes is active, respect the guard.
5. If FAIL after 2 passes: stop, summarize, ask. Don't auto-escalate to a different generator without user input.

Output format to user

When /go-loop completes, print a compact summary (no headers, no celebration):

shot-06 GOOD 0.42 (1 regen)
shot-08 PASS 0.31 (0 regens)
shot-11 FAIL — 2 regens, character metric stuck at 0.48
composite: runs/composites/2026-04-16.png
commit: a7b3c9d
notified

Related Skills

• verify — the underlying quality gate
• clip-feedback — for single-clip intent analysis
• nb2-batch — for batch-generating reference grids
• lockscenes — for locking approved items during assembly