Align All Parallel

If neither --ult nor --ust is given, run both.

Model

This skill is coordination only — run it as haiku. The alignment subagents require linguistic judgment — spawn each with model: "sonnet".

Step 1: Check Verse Count

Before spawning agents, count the verses in the chapter using mcp__workspace-tools__read_usfm_chapter on data/hebrew_bible/XX-BOOK.usfm. Count the number of \v markers in the result.

• If verse count ≤ 18: proceed to Step 2a (single batch)
• If verse count > 18: proceed to Step 2b (split into batches of 18)

Step 2a: Single Batch (≤ 18 verses)

Spawn agents in parallel:

• If running ULT: spawn ult-align subagent (model: "sonnet", skill: ULT-alignment)
• If running UST: spawn ust-align subagent (model: "sonnet", skill: UST-alignment)

Wait for both to complete. Report results.

Step 2b: Split into Batches (> 18 verses)

Determine the number of batches, then split evenly:

• Number of batches = ceil(N / 18)
• Batch size = ceil(N / numBatches) — distributes verses evenly
• Batch 1: verses 1–batchSize, Batch 2: verses (batchSize+1)–(2×batchSize), etc.
• Last batch gets the remainder

Examples:

• 22 verses → 2 batches → 11 each (v1–11, v12–22)
• 56 verses → 4 batches → 14 each (v1–14, v15–28, v29–42, v43–56)
• 176 verses → 10 batches → 18 each (v1–18, v19–36, ... v163–176)

Step 2c: Spawn ALL Batches in Parallel

Launch all batch subagents in a single message — do not wait between batches:

• For each batch K (1..numBatches):
  • If running ULT: spawn ult-align-K subagent (model: "sonnet", skill: ULT-alignment) for BOOK CH --verses START-END
  • If running UST: spawn ust-align-K subagent (model: "sonnet", skill: UST-alignment) for BOOK CH --verses START-END
• All subagents launch at once (e.g., 4 batches × 2 types = 8 parallel subagents)

Parallel cap: If the chapter has more than 5 batches per type (>90 verses), split into waves of 5 batches each. Run wave 1 (batches 1–5), wait for completion, then run wave 2 (batches 6–N). This keeps total parallel agents manageable.

Wait for all subagents to complete before proceeding.

Step 2d: Consistency Check

After all alignment batches complete, spawn a consistency checker to catch cross-batch inconsistencies. The same Hebrew word (same Strong's number) should get the same English alignment across the chapter.

For each type being aligned (ULT and/or UST), spawn a Task subagent (model: "sonnet"):

Task: "Check alignment consistency for BOOK CH (ULT|UST)"

The checker:

1. Read all partial aligned USFM files for that type
2. Extract every x-strong + x-content pairing and which English word(s) it maps to
3. Identify inconsistencies — same Strong's/content combo aligned to different English words across batches (e.g., H2617 x-content="חֶסֶד" → "faithfulness" in batch 1 but "loyalty" in batch 3)
4. Fix by choosing the most frequent alignment and updating the minority files using Edit
5. Report what was changed (or confirm no inconsistencies found)

If running both ULT and UST, spawn both checkers in parallel.

Note: The checker does NOT re-align from scratch — it only patches inconsistencies in the already-generated partial files. Skip this step for chapters with only 2 batches (≤36 verses), where inconsistency risk is low.

Step 2e: Merge

After consistency check completes, use mcp__workspace-tools__merge_aligned_usfm to assemble the full chapter:

• ULT (if applicable): call with parts = ordered array of all ULT partial files, output = output/AI-ULT/BOOK/BOOK-CH-aligned.usfm
• UST (if applicable): call with parts = ordered array of all UST partial files, output = output/AI-UST/BOOK/BOOK-CH-aligned.usfm

Do NOT use Bash, sub-agents, or manual Read+Write for the merge — use the MCP tool.

Output

• output/AI-ULT/<BOOK>/<BOOK>-<CH>-aligned.usfm - ULT with word-level alignments (if --ult or both)
• output/AI-UST/<BOOK>/<BOOK>-<CH>-aligned.usfm - UST with phrase-level alignments (if --ust or both)

Both are ready for insertion into Door43 repos.

Usage

/align-all-parallel psa 124
/align-all-parallel lam 4 --ust
/align-all-parallel lam 3 --ult --ust