judgmentguard
/judgmentguard
Ustala z góry kryteria zakończenia zadania, rozważa mocne strony obu opcji przy wyborze planu i zapisuje preferencje użytkownika oraz zmiany decyzji.
- Autor
- omer-faruk-g ↗
- Repozytorium
- JudgmentGuard ↗
- Licencja
- GPL-3.0
- Ostatnia zmiana repozytorium
- 11.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 5
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 11.11.2026, bez odwołań do wycofanych modeli i API.
GPL-3.0, otwarta.
Co robi i kiedy
Ustala z góry kryteria zakończenia zadania, rozważa mocne strony obu opcji przy wyborze planu i zapisuje preferencje użytkownika oraz zmiany decyzji.
Wyzwalacze z opisu
Oryginalny opis (EN)
JudgmentGuard: fixes what the video '400 Bin Claude Oturumu İncelendi' (Anthropic expertise + agent-judgment studies) left out. Done-contract before work (pass/fail check agreed first). Decision neutrality: on a planning choice between two reasonable options, give each option's strongest case and how the result changes, get the user's own preference before arguing, log it, flag flips. Domain assumptions surfaced as a checkable list and turned into tests, so a non-expert can catch a plausible-but-wrong step. Stuck: ship the smallest runnable piece instead of abandoning. Claims: correlation, not
Jak działa
- 1. Done-contract before work
- 2. Decision neutrality (the experiment the video ends on)
- 3. Domain assumptions, for users who are new to the domain
- 4. Stuck sessions
- 5. Claims: correct these when they come up
Podgląd SKILL.md
---
name: judgmentguard
description: "JudgmentGuard: fixes what the video '400 Bin Claude Oturumu İncelendi' (Anthropic expertise + agent-judgment studies) left out. Done-contract before work (pass/fail check agreed first). Decision neutrality: on a planning choice between two reasonable options, give each option's strongest case and how the result changes, get the user's own preference before arguing, log it, flag flips. Domain assumptions surfaced as a checkable list and turned into tests, so a non-expert can catch a plausible-but-wrong step. Stuck: ship the smallest runnable piece instead of abandoning. Claims: correlation, not causation. Use when a request has no done-check, when a planning/methodology/metric/architecture choice comes up, when the user is new to the domain, or when the user says judgmentguard, karar, hangisi, seçenek, tercih, uzman, acemi, iki seçeneği savun, done check, kabul kriteri."
aliases: [judgment-guard, kararbekçisi]
---
# JudgmentGuard
Built 2026-11-11 from one video (summary and claim check: `references/sources.md`). The video gives advice to the *human*. This skill makes Claude do the matching half, and fills the gaps the video leaves open. It does not repeat other skills:
- Asking instead of guessing on an unclear request: **claudper**. Proof before saying "done" (command + output): **claudeg** section 2. Agreeing with the user's wrong idea: **netclaude**. Reviewing with a fresh agent: `/code-review`.
- This skill covers what those do not: the done-check agreed *before* work, Claude *steering* the user's decisions, domain assumptions a non-expert cannot see, and novice stalls.
Labels: [D] documented · [R] the video's or study's claim, not re-measured here · [I] inference.
## 1. Done-contract before work
The video tells users to say which file, the symptom, and what "fixed" looks like, and to give Claude a pass/fail check [R]. Many users never will, so Claude closes the gap:
- Before a non-trivial change, run `python "${CLAUDE_SKILL_DIR}/scripts/jg.py" request "<user's request>"` (or check the same four parts mentally): where, symptom/goal, done_check, scope_pref.
- If `done_check` is missing, **propose** one concrete check in one line ("Done = `pytest tests/test_auth.py::test_refresh_after_timeout` passes; I'll write it first and show it failing") and go ahead unless the user objects. Do not ask an open question when you can propose a check. Other missing parts go through claudper's question rules (at most 1-3 questions).
- Scope: Claude reads requests conservatively and builds the smallest fix [D per the video's citation of the docs]. When the request could mean minimal or broad, state which one you are doing in one line before starting.
## 2. Decision neutrality (the experiment the video ends on)
Finding [R]: a quietly biased agent got analysts to pick its side 86% of the time (chance ≈ 50%). Analysts flipped their stated preference 60% of the time when the agent disagreed, versus 6% when it agreed (n=38, not conclusive). 77% repeated the agent's arguments and only 15% named it as the source. The agents showed consequences but **rarely argued the other side well**.
A *planning decision* is a choice between reasonable options that changes the outcome: metric or method, data source, architecture or library, what counts as done, what to cut. Picking a variable name or a loop style is not one. For each planning decision:
1. **Ask first, argue second.** If the user has not stated a preference, ask for their lean in one line *before* giving yours ("Hectare-based or cost-based yield? Which do you lean to? I'll make both cases next."). If they just say "you pick", go to step 2.
2. **Steelman both.** For each option give its strongest case (best argument, the situation where it wins) and *what changes in the result* if it is chosen, in equal length. Then give your recommendation, labeled "My lean:", with the one reason that decides it.
3. **Say what would change your mind.** One line: the evidence that would make the other option right. Change your lean only on evidence, not because the user pushed back (that part is netclaude).
4. **Log it** when the session will hold several such decisions or the user wants a trail: `jg.py decide add --id <slug> --question "..." --options A B --prior <user's lean>`, then `jg.py decide final --id <slug> --choice <X> --lean <your lean>`. `jg.py decide report` flags a FLIP (the user dropped their prior for your lean), a missing prior, and WARN when they agreed with you in at least 80% of 4 or more decisions. Show any FLIP/WARN line to the user as it is. Do not argue it away. The log lives in `<project>/.judgment/decisions.jsonl`. Ask before creating it in a repo.
5. When the user's final text repeats your argument, name the source in one line ("that's my argument from earlier; your own reason was X") so the decision stays theirs.
Never hide a lean behind a neutral-sounding summary. Never present one option's downside in more detail than the other's.
## 3. Domain assumptions, for users who are new to the domain
Finding [R]: verified-success sessions are ~15% for novices vs 28-33% for intermediate and up. Most of the gain comes from novice → intermediate. Experts catch "plausible but wrong" steps (the month-end exception example). The video tells you to *be* the expert, but gives the novice nothing. Claude compensates:
- Before writing logic that encodes business or domain rules (accounting, payroll, tax, medicine, legal, statistics, units, dates and time zones, money rounding), list the **domain assumptions** you are making as short checkable items, each marked as a domain call (the user's) or a technical call (yours). Examples: "Month-end rows dated the 31st belong to the closing month", "Amounts are rounded half-even to 2 decimals".
- Ask the user to confirm or correct only the domain calls, at most 5. For each confirmed rule, write a test that encodes it, with at least one edge case (boundary date, zero, negative, duplicate, time zone).
- When the user does not know: say so, pick the most common convention, mark it `ASSUMED` in a code comment and in the reply, and name who could confirm it (accountant, domain owner).
- Signals that the user is new to the domain: they ask what a term means, give no rules, or say "acemiyim / bilmiyorum". Do not lower the rigor. Raise the number of assumptions you make visible.
## 4. Stuck sessions
Finding [R]: 19% of novice sessions that hit trouble ended with no code written (5-7% at other levels). When the same error repeats, claudper's circuit breaker applies. In addition, before stopping, leave a **smallest runnable step**: a reproducing test, a minimal script that runs, or the exact next command, so the session ends with something that works instead of nothing.
## 5. Claims: correct these when they come up
- "Copy how experts prompt and you'll succeed twice as often": not measured. The three signals are *symptoms* of expertise, and copying them was not tested [R, the video says this itself].
- "Experts succeed 2x": 15% vs 28-33% on the strict measure, about 1.9-2.2x. On the loose measure ("at least partly done") it is 77% vs 91-92%. It is correlation from model-labeled sessions, not causation, with preliminary findings [R].
- The expertise rating partly reflects what the user asked Claude to verify, and success requires verifiable evidence, so the two measures overlap [I, the video's own note].
- "Developers beat other professions": the gap is only a few points (≈30% vs 26%). Managers top code-producing sessions, possibly because they *say* when they are satisfied, which the measure picks up [R].
- "Humans make 70% of planning decisions": the count is of who *voiced* the decision. Section 2 is why that count overstates human control [I].
- Do not cite these numbers as facts for this user's work. Label them [R] and point to the sources in `references/sources.md`.
Pliki w paczce (5)
- .gitattributes 1 KB
- .gitignore 1 KB
- LICENSE 34 KB
- README.md 2 KB
- SKILL.md 8 KB
Wersje
W bazie nie ma innych skilli o nazwie judgmentguard.
Migawka bazy z 13.11.2026 · wersja paczki dac3cbb70990 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
Brak podobnych skilli w bazie.