メニュー
End-to-end source processing -- seed, reduce, process all claims through reflect/revisit/review, archive. The full pipeline in one command. Triggers on "/pipeline", "/pipeline [file]", "process this end to end", "full pipeline".
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/FDU-INS/Insurance-Skills --skill pipeline-5コマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Complete architecture reference for the A3 insurance platform — file locations, data flow, conventions, microservice map, and fullstack patterns
Drafts U.S. commercial real estate access and indemnity (right-of-entry) agreements for pre-closing due diligence. Covers license grants, non-invasive vs invasive testing gates, insurance/endorsement requirements, indemnity with discovery carve-outs, restoration and lien remedies, confidentiality, and anti-indemnity guardrails. Trigger: access agreement, right of entry, due diligence access, Phase I/II, invasive testing, pre-PSA site inspection.
Guide actuarial analysis for insurance pricing. TRIGGERS - Use when user needs help with actuarial-analysis related tasks.
SOC 職業分類に基づく
SKILL.md を表示中
| name | pipeline-5 |
| description | End-to-end source processing -- seed, reduce, process all claims through reflect/revisit/review, archive. The full pipeline in one command. Triggers on "/pipeline", "/pipeline [file]", "process this end to end", "full pipeline". |
| version | 1.0 |
| generated_from | arscontexta-v1.6 |
| user-invocable | true |
| context | fork |
| model | opus |
| allowed-tools | Read, Write, Edit, Grep, Glob, Bash, Task |
| argument-hint | [file] — path to source file to process end-to-end |
Target: $ARGUMENTS
Parse immediately:
--handoff: output RALPH HANDOFF block at end (for chaining)Read arscontexta/ops/derivation-manifest.md (or fall back to arscontexta/ops/derivation.md) for domain vocabulary mapping. All output must use domain-native terms. If neither file exists, use universal terms.
START NOW. Run the full pipeline.
The pipeline chains four phases. Each phase uses skill invocation or /ralph for subagent-based processing. State lives in the queue file — the pipeline is stateless orchestration on top of stateful queue entries.
Source file
|
v
Phase 1: /seed — create extract task, move source to archive
|
v
Phase 2: /extract (via /ralph) — extract claims from source
|
v
Phase 3: /ralph (all claims) — create -> reflect -> reweave -> verify
|
v
Phase 4: /archive-batch — move task files, generate summary
|
v
Complete
The pipeline is the convenience wrapper. /ralph is the engine. /seed is the entry point.
Invoke /seed on the target file to create the extract task, check for duplicates, and move the source to its archive folder.
How to invoke:
Use the Skill tool if available, otherwise execute the /seed workflow directly:
Capture from seed output:
Report: $ Seeded: {source-name}
If seed reports the file was already processed: Ask the user whether to proceed or skip. Do NOT auto-skip — the user may want to re-process with different scope.
Process the extract task via /ralph. This spawns a subagent that runs /extract, extracting claims from the source and creating task entries in the queue.
How to invoke:
/ralph 1 --batch {batch_id} --type extract
Or via Task tool:
Task(
prompt = "Run /ralph 1 --batch {batch_id} --type extract",
description = "extract: {batch_id}"
)
After completion, read the queue to count extracted claims and enrichments:
Check how many pending tasks exist for this batch. The reduce phase creates 1 queue entry per claim and 1 per enrichment.
Report:
$ Extracted: {N} notes, {M} enrichments
Processing {total_tasks} tasks through the pipeline...
If zero claims extracted: Report the issue. For TFT sources, zero extraction is a bug — the source almost certainly contains extractable content. Ask the user whether to retry with different scope or skip.
Count total pending tasks for this batch from the queue. Then process all of them through the full phase sequence.
How to invoke:
/ralph {remaining_count} --batch {batch_id}
Or via Task tool:
Task(
prompt = "Run /ralph {remaining_count} --batch {batch_id}",
description = "process: {batch_id} ({remaining_count} tasks)"
)
This processes every claim through: create -> reflect -> reweave -> verify. And every enrichment through: enrich -> reflect -> reweave -> verify.
Each phase runs in an isolated subagent with fresh context. /ralph handles all the orchestration: subagent spawning, handoff parsing, queue advancement, learnings capture.
Progress reporting:
The /ralph invocation reports progress per task. The pipeline relays this:
$ Processing note 1/{total}: {title}
$ create... done
$ reflect... done (3 connections found)
$ reweave... done (2 notes updated)
$ verify... done (PASS)
For large batches (20+ claims): /ralph handles context isolation automatically via subagents. The pipeline does NOT need to chunk — /ralph processes N tasks sequentially with fresh context per phase.
After /ralph finishes, verify all tasks for this batch are done.
Check the queue: count tasks for this batch that are NOT done.
If tasks remain pending:
/ralph --batch {batch_id} to continue from where it stopped"If all tasks are done: Proceed to Phase 5.
When all tasks for the batch are complete, archive the batch.
How to invoke:
/archive-batch {batch_id}
Or execute directly:
arscontexta/ops/queue/ to arscontexta/ops/queue/archive/{date}-{batch_id}/{batch_id}-summary.mdThe summary should include:
--=={ pipeline }==--
Source: {source_file}
Batch: {batch_id}
Extraction:
notes extracted: {N}
Enrichments identified: {M}
Processing:
notes created: {N}
Existing notes enriched: {M}
Connections added: {C}
indexs updated: {T}
Older notes updated via reweave: {R}
Quality:
All verify checks: {PASS/FAIL count}
Archive: arscontexta/ops/queue/archive/{date}-{batch_id}/
Summary: {batch_id}-summary.md
notes created:
- [[claim title 1]]
- [[claim title 2]]
- ...
If --handoff flag was set, also output:
=== RALPH HANDOFF: pipeline ===
Target: {source_file}
Work Done:
- Seeded source: {batch_id}
- Extracted {N} notes and {M} enrichments
- Processed all claims through 4-phase pipeline
- Archived batch to {archive_path}
Files Modified:
- docs/notes/ ({N} new notes)
- arscontexta/ops/queue/archive/{date}-{batch_id}/ (archived)
Learnings:
- [Friction]: {description} | NONE
- [Surprise]: {description} | NONE
- [Methodology]: {description} | NONE
- [Process gap]: {description} | NONE
Queue Updates:
- All tasks for batch {batch_id} marked done and archived
=== END HANDOFF ===
Phase failure at any stage:
/ralph --batch {batch_id} to continue from where it stopped"The pipeline is resumable. Queue state persists across sessions:
Seed failure: If /seed fails (file not found, duplicate detected and user declines), stop the pipeline entirely.
Extract failure: If /extract extracts zero claims, report and stop. Do not proceed to an empty processing phase.
Processing failure: If /ralph fails mid-batch, the queue preserves state. Individual claims resume from their failed phase on next /ralph invocation.
Archive failure: If archiving fails, the claims are still created and connected. Only the organizational cleanup is missing — re-run /archive-batch manually.
The pipeline is designed to be interrupted and resumed at any point:
| Interrupted At | How to Resume |
|---|---|
| Before seed | Run /pipeline again (starts fresh) |
| After seed, before reduce | /ralph 1 --batch {id} --type extract |
| After reduce, during claims | /ralph --batch {id} (picks up from failed phase) |
| After all claims, before archive | /archive-batch {id} |
State lives in the queue file. The pipeline reads queue state, not session state. This means you can interrupt, close the session, and resume later.
No target file: List arscontexta/inbox/ candidates, suggest the best one based on age and relevance.
Source already seeded: /seed detects this and asks the user. If they decline, the pipeline stops cleanly.
Large source (2500+ lines): /extract handles chunking automatically. The pipeline does not need special handling.
No arscontexta/ops/derivation-manifest.md: Use universal vocabulary for all output.
never:
always: