writing-th-learn

name: writing-th-learn description: Extracts Thai writing patterns by comparing draft vs edited versions. links learnings to the session writing plan + example report style reference.

/writing-th-learn

Learning-only companion for /writing-th. Compares a Thai draft and its human-edited version, then stores what changed as reusable patterns.

When to use this skill

• After you used /writing-th to produce a draft and you have an edited version (...-edited.md).
• When closing a session with /rrr and you want to capture Thai writing style learnings separate from the general retrospective.
• When you have any Thai draft/edited pair and want to update writing-th's understanding of your word choice and phrasing.

When NOT to use this skill

• If there is no edited version yet (only a draft exists) — use /writing-th first to draft and let the human edit.
• If the change is not about Thai wording, phrasing, or narrative structure (for example pure data or numbers updates).
• If you only need to generate a new Thai draft — use /writing-th instead.

Inputs required

1. Mode: report or article (must match the mode originally used in /writing-th).
2. Writing plan path (recommended): the anchor plan markdown file for the session.
3. Draft file path: markdown file.
4. Edited file path: markdown file.
5. (Optional) Example report path: the style reference used in the session (if any).
6. (Optional) Context tags: short tags for oracle_learn (e.g., thai-tone, gov-report, interview-prep).

Naming rule (recommended, but not mandatory)

• Preferred: ...-draft.md and ...-edited.md pairing.
• If real files do not conform, proceed only if a clear mapping is recorded (Nothing is Deleted):
  • Add a small mapping block in the writing plan (recommended template is in /writing-th).

Workflow

1. Validate inputs

   • Confirm that both files exist and are markdown.
   • Prefer: they share the same base name and differ only by -draft vs -edited.
   • If not: require explicit mapping (in the plan or user message) before continuing.

2. Read session context (Option C)

   • If a writing plan path is provided:
     • Read it and find the latest “Session Style Pack Summary”.
     • Capture:
       • example report reference + permission line
       • session terminology / flow constraints
       • safety rails (citations/hedging/no hallucinations)

3. Read both files

   • Read the full content of the draft and edited files.

4. Compare and extract patterns

   • Focus on what the human changed, not on line-by-line diffs.
   • Identify and summarize:
     • Word choice and phrasing:
       • recurrent substitutions (e.g., more formal vs informal verbs, consistent noun choices);
       • changes in tone (more/less formal, softer/stronger hedging);
       • preference for certain connectors, transitions, or emphasis styles.
     • Semantic arrangement and structure:
       • reordering of sentences or paragraphs to improve flow;
       • patterns in how the human introduces context, states key messages, and provides evidence;
       • ways the human simplifies or elaborates concepts for clarity.
   • Keep the summary concise and pattern-based (no full text copies).

5. Write learning note to ψ/memory

   • Compose a learning note in Thai (with English only where necessary) with sections:

     • ## Word choice and phrasing
     • ## Semantic and structure patterns
     • ## Implications for future drafts
     • ## Session context (traceability)

   • Save it to:

     • ψ/memory/learnings/YYYY-MM-DD_writing-th-<mode>-learn.md where <mode> is report or article.

   • If multiple learning notes already exist for the same day and mode, append a new section with a clear heading instead of overwriting.

   • In ## Session context (traceability) include links:

     • writing plan path (if any)
     • draft path + edited path
     • example report path (if any)

6. oracle_learn call

   • Call oracle_learn with:
     • pattern: the distilled patterns from the learning note (especially "Implications for future drafts").
     • concepts: tags including writing-th, the mode (report or article), and any additional context tags.
     • source: a short identifier like "writing-th-learn: draft-vs-edited".

   Materialization guardrail (IMPORTANT)

   • After oracle_learn(), immediately read the returned file path.
   • If read fails with ENOENT, materialize the file manually (patch tooling), per:
     • ψ/memory/learnings/2026-03-27_oracle-learn-file-materialization-guardrail.md

7. Link back to execution skill

   • In the learning note, add a short line that this file is discoverable by /writing-th via MCP search:

     • e.g. "เรียนรู้จากการเปรียบเทียบฉบับร่างและฉบับแก้ไขสำหรับ writing-th (mode: report/article)".

   • If a writing plan exists, append an “Learn-back recorded” line into the plan (append-only) with a link to the learning note.

8. Optional promotion to resonance (with explicit human approval)

• Useได้เมื่อ:

• • pattern เดิม (เช่น วิธีตั้งประโยค วิธีจัดลำดับบริบท → ประเด็นหลัก → นัยเชิงยุทธศาสตร์) ปรากฏซ้ำใน อย่างน้อย 2 session และ
• • คุณในฐานะมนุษย์ตัดสินใจแล้วว่าอยากให้ pattern นั้นกลายเป็น “กติกาถาวร” ของการเขียน (ไม่ใช่แค่ preference ชั่วคราวของโปรเจกต์เดียว)

• ขั้นตอน:

• 1. เปิด learning note ล่าสุดที่เกี่ยวข้อง (ใน ψ/memory/learnings/YYYY-MM-DD_writing-th-<mode>-learn.md) และเลือกเฉพาะ pattern ที่ควรโปรโมต
• 2. สร้างบล็อกข้อความสั้น ๆ ในรูปแบบ:

• ```markdown
  

• ## YYYY-MM-DD — promoted from writing-th-learn (mode: report/article)
  

• - บรรยาย pattern เชิงกติกาแบบสั้น กระชับ และทั่วไปพอใช้ได้ข้ามโปรเจกต์
  

• - ระบุว่าเรียนรู้จาก draft/edited คู่ใด (ลิงก์ path แบบย่อ)
  

• ```
  

• 3. เพิ่มบล็อกนี้แบบ append-only ลงใน

• - [`ψ/memory/resonance/writing-style-th.md`](ψ/memory/resonance/writing-style-th.md)
  

• - หรือ [`ψ/memory/resonance/citation-style-th.md`](ψ/memory/resonance/citation-style-th.md) ถ้าเป็นเรื่อง citation โดยเฉพาะ
  

• +- Guardrail:
• • ห้ามลบบรรทัดเก่าใน resonance (Nothing is Deleted)
• • ต้องมี “เหตุผล” ที่ชัดเจนว่าทำไม pattern นี้จึงควรใช้ข้ามโปรเจกต์ ไม่ใช่ข้อยกเว้นเฉพาะกรณี
• • การตัดสินใจโปรโมตยังคงต้องอาศัยการยืนยันของมนุษย์ ไม่ใช่ agent ตัดสินใจเองลำพัง

Notes for /rrr integration

• /rrr should not recreate this logic; instead, it can:
  • Detect when a Thai draft/edited pair was created in the session; then
  • Invoke /writing-th-learn once per pair to record learnings.
• Retrospective and Thai-writing learnings must stay in separate files:
  • /rrr writes session retrospectives under ψ/memory/retrospectives/... and general learnings under ψ/memory/learnings/....
  • /writing-th-learn writes Thai-writing-focused learnings with the writing-th-<mode>-learn suffix so they are easy to query later.