Lofn Step 11 Packager: Full-Context Refinement Prompts

Where this fits

You've just finished a Lofn music run—step 10 produced pairs, and now the final refinement pass (Step 11) needs a high-capability model like Codex or Opus to polish each track. But copying the personality YAML, the Suno v5.5 rules, golden songs, run context, and prior drafts into a coherent 150KB prompt is tedious, error-prone, and eats an hour per pair. The lofn-step11-packager skill assembles that full bundle into a single paste-ready file per pair, so you can move from run completion to refinement without manual copy-paste gymnastics.

Why agents benefit

• Reduces context-assembly mistakes: The skill pulls the exact files it knows are canonical, so the refinement model gets the same golden-output quarantine and no-skip rules every time—no missing sections.
• Saves significant token/time: Instead of asking the agent to read and combine 10+ files, it emits one well-structured prompt, making the whole refinement loop faster and more predictable.
• Enforces policy automatically: It binds to the .claude skill's authority and execution deltas, so agents don't have to remember version-specific rules (like judge separation or publish bar)—they're baked into the prompt.
• Works for any step-11 demand: Whether you need one pair or all pairs, the skill scales the bundle generation to the requested scope.
• Keeps refinement reproducible: Each prompt contains full run context, so even a fresh model instance can pick up where the last one left off without asking clarifying questions.

Practical scenarios

Shipping a finished run to Codex

You've completed a 12-pair Lofn run and need to refine the final versions. Instead of manually constructing a 180KB prompt for each pair, you ask the agent to "package step 11 for all pairs". The skill generates 12 separate prompt files, each with the full personality, golden songs, the pair's step-10 output, and the canonical refinements—ready to paste into Codex.

Auditing refinement quality for a specific pair

A particular pair's output sounds off. You want to compare it against the golden standard. The packager can create a focused bundle for just that pair, including the step-10 draft and any prior step-11 revision, so you can trace exactly what the refinement model saw and where it might have gone wrong.

Handing off to a new agent or contractor

You've switched tools or need a second opinion. The packaged prompt contains everything the new model needs to understand the entire run's context—no internal assumptions. This makes it a great handoff artifact for external reviewers or for reopening a run weeks later.

Add it to your agent workflow

Install with:

npx skills add LocalSymmetry/lofn --skill lofn-step11-packager

Then, in your agent, simply ask:

"Package step 11 for pair 4 and 7."

The skill will produce two full-context refinement prompts, each ~120–200KB, with all required sections. You can inspect them before pasting into your preferred refinement model. For a deeper look at how this skill fits the wider ecosystem, check its OpenAgentSkill listing.

Compare before adopting

• Quality signals: Look at how often the repo updates (the authority date is 2026-07-01—recent), whether the skill's output matches your proven refinement templates, and if the canonical structure is actually used in your workflow.
• Maintenance freshness: The skill mirrors a .claude twin; ensure that one is actively maintained so the Codex mirror stays in sync.
• Alternatives: You could write your own prompt-assembly script or rely on a generic file-combiner, but that sacrifices the Lofn-specific rules and golden-output handling.
• Workflow fit: If you don't use Lofn's step 10/11 pipeline, this skill is irrelevant. For heavy Lofn users, it removes a major manual bottleneck.

Why it is worth tracking

With 22 stars and a clear, narrow purpose, this skill isn't a generic tool—it's a workflow multiplier for a specific, complex pipeline. It demonstrates a pattern of encoding domain-specific rules into a packager, which is increasingly valuable as agent workflows get more specialized. Track it if you're using Lofn for music generation or if you're building skills that assemble complex prompts; it's a good example of how to turn a manual, error-prone step into a one-command operation.