GPT Review — P10B-1A-HASH Agent Prompt v0.1

Date: 2026-04-29

Verdict

PASS with 3 minor patches before dispatch.

The prompt is directionally correct and solves the SHA/vocab evidence gap without redesigning the D32 segmentation. It is a proper read-only micro-step.

What is good

• Keeps the 23-unit segmentation stable.
• Does not allow insert/update/delete/DDL.
• Computes full source SHA and per-unit body SHA before insert/package.
• Queries production vocab rather than relying only on seed evidence.
• Adds body coverage verification.
• Uploads an evidence report and stops.

Required patches

Patch 1 — Do not require blanket git commit

The prompt says:

Git commit cuối prompt (VPS = code SSOT Đ41)

This is too broad for a read-only evidence task where canonical output is Agent Data KB. Use the established conditional wording:

• run git status in any repo touched;
• commit only if persistent tracked repo files were intentionally created/modified;
• if only temp files and KB report are produced, no git commit is required;
• do not create files solely to satisfy commit.

Patch 2 — Make body coverage precise, not vague

The prompt says 0 content gap, 0 unexpected overlap, but extraction excludes headings and --- separators.

Patch wording:

• coverage must list exactly what is excluded:
  • root title;
  • markdown heading markers that are represented as titles;
  • --- separators;
  • intentional container placeholders;
• leaf unit bodies should cover all substantive body text;
• report excluded text counts and samples;
• unexpected gap/overlap = FAIL.

Patch 3 — Vocab active check must adapt to actual columns

The query currently omits lifecycle/status filtering. Patch:

• inspect tac_section_type_vocab columns first;
• if lifecycle_status exists, require lifecycle_status='active';
• if an active boolean exists, require active=true;
• otherwise report the available status columns and verify code existence.

Expected 8 codes must be returned and usable.

Non-blocking note

sha256sum of body depends on exact newline normalization. Agent should specify whether body hash uses exact extracted UTF-8 bytes after source-slice extraction, preserving newlines as in source. Use the same convention later in P10B-1B.

Directive to Opus 4.6

Patch the P10B-1A-HASH prompt with the 3 items above. No need for another long review if patched cleanly.

After patch, mark as:

GPT AUTHORIZED — execute immediately

and dispatch to Agent.

Status

• P10B-1A-HASH prompt: PASS with minor patch.
• Next: patch and dispatch Agent read-only micro-step.