workflow-authoring

Workflow Authoring

Given a cluster id and 2-3 sample session paths, produce a polished .workflow.js (Workflow tier) or SKILL.md (Skill tier) at /tmp/sublimate-run/cluster-<id>.workflow.js or /tmp/sublimate-run/cluster-<id>/SKILL.md.

Inputs

• cluster_id — string
• tier — WORKFLOW or SKILL
• samples — array of 2-3 absolute paths to .jsonl session files
• skeleton_path — /tmp/sublimate-run/cluster-<id>.workflow.js (PrefixSpan raw output)

Procedure

1. Read samples. Open every sample path. Focus on user messages, tool sequences, and final assistant turns. The tool sequence alone is insufficient — read what the user actually wanted accomplished.
2. Name. Choose meta.name as a short verb-noun (generate-cover, audit-fix-verify, mine-weekly-retro). Avoid generics like do-thing.
3. Phase. Group tool calls into goal-oriented phases. Each phase has one goal — audit, fix, verify, emit, etc. Boundaries are determined by goal change, not by tool change.
4. Author prompts. For each phase write a coherent agent({prompt}) that generalizes — not a verbatim replay of the sample. The polished prompt should work for the next run, not just past runs.
5. Fan-out. Any phase with N independent units gets parallel(). Default to serial; only parallelize when independence is provable from samples.
6. Deterministic only. No wall-clock (Date.now), no RNG (Math.random), no new Date(). Use meta.id plus inputs for any keying.

Output

Overwrite the skeleton:

• Workflow tier → /tmp/sublimate-run/cluster-<id>.workflow.js
• Skill tier → /tmp/sublimate-run/cluster-<id>/SKILL.md

Return a short report:

output: <path>
summary: <one line describing what the workflow/skill does>
samples_read: [<path1>, <path2>, ...]