docs(#1578): rebase B2+B3 agent prompt changes

This commit is contained in:
Alex V.
2026-06-26 11:41:31 +03:00
parent a467dad065
commit f15c25867d
25 changed files with 348 additions and 54 deletions

View File

@@ -20,6 +20,104 @@ If the prompt contains a `<required_reading>` block, use the `Read` tool to load
@~/.claude/gsd-core/references/untrusted-input-boundary.md
<extraction_discipline>
This is **rule-application, not generation.** Apply the taxonomy / precedence rules directly to what the source actually contains. Do not infer, embellish, summarize creatively, or add any content not present in the source. Output only the required structure; when the source is silent on a field, mark it absent rather than guessing. (2505.11423 — applies here as a simple mechanical constraint: mark absent rather than fabricate.)
</extraction_discipline>
<few_shot_exemplars>
These worked examples show the exact input→output contract. Apply the same pattern to new inputs.
**Exemplar 1 — Clean ADR case**
Input: file `docs/adr/0003-choose-postgres.md`, first 50 lines contain:
```
---
status: Accepted
---
# ADR-0003 Use PostgreSQL as primary datastore
## Context
We evaluated SQLite, MySQL, and Postgres. Team has prior Postgres expertise.
## Decision
Use PostgreSQL 15+ for all relational data.
## Consequences
Operators must provision a Postgres instance.
```
Output:
```json
{
"source_path": "docs/adr/0003-choose-postgres.md",
"type": "ADR",
"confidence": "high",
"manifest_override": false,
"title": "ADR-0003 Use PostgreSQL as primary datastore",
"summary": "Chose PostgreSQL 15+ as the primary relational datastore based on team expertise.",
"scope": ["PostgreSQL", "primary datastore", "relational data"],
"cross_refs": [],
"locked": true,
"precedence": null,
"notes": ""
}
```
**Exemplar 2 — Ambiguous / UNKNOWN case**
Input: file `docs/notes/meeting-2024-01-15.md`, first 50 lines contain:
```
# Meeting notes Jan 15
Discussed caching options. No decision reached.
Bob mentioned Redis might be faster. Action: revisit next week.
```
Output:
```json
{
"source_path": "docs/notes/meeting-2024-01-15.md",
"type": "UNKNOWN",
"confidence": "low",
"manifest_override": false,
"title": "Meeting notes Jan 15",
"summary": "Meeting notes discussing caching options; no decision or requirement recorded.",
"scope": ["caching", "Redis"],
"cross_refs": [],
"locked": false,
"precedence": null,
"notes": "No ADR/PRD/SPEC signals. Prose-only meeting notes. No status field, no decision statement, no requirements. Mark as UNKNOWN — user must type-tag via manifest."
}
```
**Exemplar 3 — Edge case: PRD with ADR-like section header**
Input: file `docs/prd/user-auth.md`, contains:
```
# User Authentication PRD
## User Stories
As a user, I want to log in with email+password.
## Acceptance Criteria
- Login succeeds with valid credentials
- Login fails with wrong password (error message shown)
## Decision
Use JWT tokens. (inherited from ADR-0005)
```
Output (note: one `## Decision` section inherited from ADR does NOT override strong PRD signals):
```json
{
"source_path": "docs/prd/user-auth.md",
"type": "PRD",
"confidence": "medium",
"manifest_override": false,
"title": "User Authentication PRD",
"summary": "Requirements for email+password login with JWT tokens.",
"scope": ["user authentication", "login", "JWT"],
"cross_refs": [],
"locked": false,
"precedence": null,
"notes": "Contains one '## Decision' section but dominant signals are user stories + acceptance criteria → PRD. ADR reference recorded in cross_refs if a link is present."
}
```
</few_shot_exemplars>
<why_this_matters>
Your classification drives extraction. If you tag a PRD as a DOC, its requirements never make it into REQUIREMENTS.md. If you tag an ADR as a PRD, its decisions lose their LOCKED status and get overridden by weaker sources. Classification fidelity is load-bearing for the entire ingest pipeline.
</why_this_matters>
@@ -111,6 +209,13 @@ Regardless of type, extract:
- **locked_markers** — for ADRs only: does status read `Accepted` (locked) vs `Proposed`/`Draft` (not locked)? Set `locked: true|false`.
</step>
<terminal_output_schema_restatement>
**Output contract reminder (2506.00069 — restate schema immediately before writing):**
You MUST write exactly one JSON object matching this schema — no extra fields, no omissions:
`{ source_path, type (ADR|PRD|SPEC|DOC|UNKNOWN), confidence (high|medium|low), manifest_override (bool), title (string), summary (≤30 words), scope (string[]), cross_refs (string[]), locked (bool), precedence (int|null), notes (string, omit if high confidence) }`
`locked: true` only for ADR with `Accepted` status. `manifest_override: true` only if MANIFEST_TYPE was provided. Fields absent in source → mark absent (empty array / empty string / false), never fabricate.
</terminal_output_schema_restatement>
<step name="write_output">
Write to `{OUTPUT_DIR}/{slug}-{source_hash}.json` where `slug` is the filename without extension (replace non-alphanumerics with `-`), and `source_hash` is the first 8 hex chars of SHA-256 of the **full source file path** (POSIX-style) so parallel classifiers never collide on sibling `README.md` files.