Zum Inhalt

Büro-/Wissensarbeit-Pack Implementation Plan

For agentic workers: REQUIRED SUB-SKILL: Use superpowers:executing-plans (inline, single session). Steps use checkbox (- [ ]) syntax for tracking.

Goal: Ship packs/buero.yaml — a second eval pack (Büro-/Wissensarbeit-Assistent) next to ndassist, pure data, with a shipped-pack parse test.

Architecture: A pack is data, not code (touchstone/pack.py validates it). The full prompt content (25 prompts, green/red flags, dimensions, K.-o.) is fixed in the approved spec docs/superpowers/specs/2026-06-27-buero-pack-design.md and the adversarially-reviewed workflow design. Implementation = author the YAML + a parse/invariant test mirroring test_shipped_ndassist_pack_parses + doc touch-ups.

Tech Stack: Python 3.12, uv, pydantic (pack validation), pytest, ruff, mypy strict, PyYAML.

Global Constraints

  • id: buero · title: "Büro- & Wissensarbeit-Assistent" · version: 1.
  • 25 prompts in 5 categories A–E (5 each); dimension weights Q1..Q7 = 3,3,2,2,2,2,1 (sum 15, max_weighted 75).
  • ko_rule: dimension: Q1, threshold: 2, red_flag_prompts: [A4, B3, C4, D1, E1, E2].
  • safety_critical: true only on A4, B3, C4, D1, E1, E2; format_strict: true only on A3, A5, B4, C5; repeats: 2 only on A4, B3, C4, D1, E1, E2, E4; max_tokens: null everywhere.
  • prompt_variants: baseline (professional office-assistant system prompt) + none (null).
  • sampling: temperature 0.0, seed 42.
  • German prompts/flags; identifiers/keys English-keyed YAML. Source text lives inside the prompt string.
  • All edits keep pytest -q, ruff check, ruff format --check, mypy touchstone/ green.

Task 1: Author packs/buero.yaml + shipped-pack parse test (TDD)

Files: - Create: packs/buero.yaml - Modify: tests/test_pack.py (append one test)

Interfaces: - Consumes: touchstone.pack.load_pack(path) -> Pack; Pack.all_prompts(), Pack.dimensions, Pack.max_weighted, Pack.ko_rule (existing). - Produces: a shipped data file at the conventional path packs/<id>.yaml, discoverable by the GUI glob and bundles.pack_rel.

  • [ ] Step 1: Write the failing test — append to tests/test_pack.py:
def test_shipped_buero_pack_parses():
    from pathlib import Path

    root = Path(__file__).resolve().parent.parent
    pack = load_pack(root / "packs" / "buero.yaml")
    # Büro brief: 25 prompts, 7 weighted dimensions summing to 15 (max 75).
    assert pack.id == "buero"
    assert len(pack.all_prompts()) == 25
    assert sum(d.weight for d in pack.dimensions) == 15
    assert pack.max_weighted == 75
    # K.-o. = no-hallucination on Q1, narrow confabulation-bait pool.
    assert pack.ko_rule.dimension == "Q1"
    assert pack.ko_rule.threshold == 2
    assert pack.ko_rule.red_flag_prompts == ["A4", "B3", "C4", "D1", "E1", "E2"]
    # Field conventions: exactly the 6 K.-o. prompts are safety_critical.
    sc = sorted(p.id for _, p in pack.all_prompts() if p.safety_critical)
    assert sc == ["A4", "B3", "C4", "D1", "E1", "E2"]
    # Every prompt has at least one green and one red flag (Judge rubric).
    assert all(p.green_flags and p.red_flags for _, p in pack.all_prompts())
  • [ ] Step 2: Run test to verify it fails

Run: uv run pytest tests/test_pack.py::test_shipped_buero_pack_parses -q Expected: FAIL — file packs/buero.yaml does not exist (load_pack raises).

  • [ ] Step 3: Author packs/buero.yaml

Write the full pack following packs/ndassist.yaml structure and the spec inventory. Content source = the approved, adversarially-reviewed workflow design (frame + 5 categories), with the three resolved decisions applied: narrow K.-o. pool (A4,B3,C4,D1,E1,E2 — D2/D3/D5 safety_critical: false), E3 = Q7 verbindliche-Zusage prompt, A3 green-flag sentence-count hardened ("Anrede/Gruß zählen nicht als Sätze; nur der Mailrumpf zählt"). Top-level keys in order: id, title, version, description, scale, dimensions, ko_rule, prompt_variants, sampling, categories.

  • [ ] Step 4: Run the new test to verify it passes

Run: uv run pytest tests/test_pack.py::test_shipped_buero_pack_parses -q Expected: PASS.

  • [ ] Step 5: Commit
git add packs/buero.yaml tests/test_pack.py
git commit -m "feat(packs): Büro-/Wissensarbeit eval pack (buero) + parse test"

Task 2: Doc touch-ups (AGENTS.md + README)

Files: - Modify: AGENTS.md (the "packs/ndassist.yaml is the first one" sentence) - Modify: README.md, README.de.md (wherever ndassist is named as the only/example pack)

Interfaces: none (prose only).

  • [ ] Step 1: Update AGENTS.md — change the qualitative-half paragraph so it names both packs (ndassist = ND-Assistent, buero = Büro-/Wissensarbeit-Assistent) instead of "the first one".

  • [ ] Step 2: Update README.md / README.de.md — wherever the eval example uses packs/ndassist.yaml, add a one-line note that packs/buero.yaml is a second shipped pack (no command rewrite needed).

  • [ ] Step 3: Verify nothing references a removed/renamed symbol

Run: git grep -n "first one\|erste" AGENTS.md Expected: no stale "first/only pack" claim remains.

  • [ ] Step 4: Commit
git add AGENTS.md README.md README.de.md
git commit -m "docs: name buero pack alongside ndassist (two shipped packs)"

Task 3: Full verification gate

Files: none (verification only).

  • [ ] Step 1: Run the full suite + lint + types

Run: 

uv run pytest -q
uv run ruff check .
uv run ruff format --check .
uv run mypy touchstone/
Expected: all green; test count = previous + 1 (the new buero test). No GUI test breaks from the added pack file.

  • [ ] Step 2: (If anything red) fix inline, re-run. Do not proceed to merge until all four are green.

Self-Review

Spec coverage: Identity, scale, 7 dimensions, K.-o. (narrow pool), 25-prompt inventory, field conventions, E3 repurpose, A3 hardening, acceptance criteria 1–5 → all carried by Task 1 (YAML+test) + Task 2 (docs) + Task 3 (verify). Acceptance #4 (GUI lists buero, no test breaks) is covered by Task 3 Step 1.

Placeholder scan: Test code is concrete. The YAML body is authored in Task 1 Step 3 from the named source (spec + approved design) rather than re-pasted here — deliberate, to avoid triplicating ~400 lines of flag text; the invariant test pins the structural contract.

Type consistency: Uses only existing Pack/load_pack API as in the neighboring test_shipped_ndassist_pack_parses. No new symbols.