visual-critic
/visual-critic
Koordynuje i ocenia zadania wizualizacji danych: prowadzi wykres przez kolejne etapy albo ocenia gotową grafikę według rubryki błędów.
- Autor
- Lmandrake ↗
- Repozytorium
- MandrakeWorkflows ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 11.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- 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.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Koordynuje i ocenia zadania wizualizacji danych: prowadzi wykres przez kolejne etapy albo ocenia gotową grafikę według rubryki błędów.
Oryginalny opis (EN)
Orchestrator and judge for the visual-reasoning stack. Use to run an end-to-end visualization task (routing through intent → encoding → composition → scientific-visualization → render-qa) OR to critique/grade an existing visual against the failure taxonomy and rubric. Invokes the other skills; it does not duplicate their rules.
Jak działa
- Purpose
- When to invoke
- Inputs
- Required intermediate representation
- Procedure
- Hard constraints
- Strong heuristics
- Context-dependent guidance
- Failure modes
- Validation checks
Podgląd SKILL.md
--- name: visual-critic description: > Orchestrator and judge for the visual-reasoning stack. Use to run an end-to-end visualization task (routing through intent → encoding → composition → scientific-visualization → render-qa) OR to critique/grade an existing visual against the failure taxonomy and rubric. Invokes the other skills; it does not duplicate their rules. --- # Visual Critic (orchestrator) ## Purpose Drive the whole reasoning sequence and adjudicate the result. The critic decides which skills to invoke and in what order, holds the shared artifact state, applies the failure taxonomy and the scoring rubric, and owns the final SHIP decision. It is the single place where the hard-vs-soft discipline (QA-01) is enforced across skills. ## When to invoke - For any nontrivial "make me a good visualization of X" request (end-to-end mode). - To review, grade, or repair an existing figure/slide/diagram (critique mode). ## Inputs - The user request and data, plus medium/audience if known. - In critique mode, the existing artifact (rendered) to evaluate. ## Required intermediate representation A **case file** that accumulates each skill's output: ``` intent_record (from visual-intent) encoding_table (from visual-encoding) fidelity_record (from scientific-visualization, if scientific) composition_plan (from visual-composition) qa_report (from visual-render-qa) failure_scan (which of the 7 levels are implicated) rubric_score (100-pt, with hard-gate result) verdict ``` ## Procedure 1. **Route.** Always start with `visual-intent`. Then `visual-encoding`. If the subject is scientific/spatial/network/diagram/map, invoke `scientific-visualization` in parallel with encoding/composition. Then `visual-composition`. Then `visual-render-qa`. Do not re-implement their rules here — call them and collect their outputs into the case file. 2. **Follow the reasoning sequence** (do not skip levels): intent → abstraction → task → representation → encoding → hierarchy → composition → annotation → style → render → inspect → repair (S01 nested-model discipline; CRIT-01/02/06). 3. **Failure scan.** Locate any defect at the correct of the seven failure levels: (1) wrong problem, (2) wrong abstraction, (3) wrong encoding, (4) wrong composition, (5) wrong scientific representation, (6) physical-rendering defect, (7) wrong communication strategy. Fixing a defect at the wrong level wastes effort (CRIT-06). 4. **Gate then score.** Take the `qa_report`. If any hard gate failed (QA-02), the verdict is REPAIR regardless of rubric score — route the prescribed fix to the owning skill and re-run render-qa. Only once all hard gates pass, compute the soft rubric score to choose among valid alternatives (QA-01/03). 5. **Justify** each nontrivial choice against a principle or evidence (CRIT-01), and when presenting a result as good, say what makes it work (CRIT-03). 6. **Escalate** genuinely context-dependent decisions (UNC-01 baseline; contested rules per CRIT-04) to the user rather than resolving silently. 7. **Verdict:** SHIP / REPAIR (with the level + owning skill named) / ESCALATE. ### Rubric (100 pts; soft ranking only — hard gates are pass/fail and override) Semantic fidelity 25 · Scientific/quantitative integrity 20 · Perceptual effectiveness 15 · Layout/render integrity 15 · Communication hierarchy 10 · Accessibility 5 · Medium fitness 5 · Aesthetic coherence 5. A hard-gate failure (see `visual-render-qa` step 2) blocks SHIP no matter the score. ## Hard constraints - QA-01, QA-02 (MUST): validity gate precedes and overrides scoring. - CRIT-02 (MUST): the evaluation method must match the level of the claim. - CRIT-06 (AVOID): do not fixate on encoding while ignoring upstream levels. - The critic MUST NOT ship an artifact whose `qa_report` verdict is not SHIP. ## Strong heuristics - CRIT-01, CRIT-03, CRIT-05 (SHOULD): justify choices, explain what works, subordinate aesthetics to understanding. ## Context-dependent guidance - CRIT-04 (MAY): model contested rules (axis-zero off-bars) as weighted, not absolute; expose the trade-off. - In critique mode with no editable source, downgrade REPAIR actions to prioritized recommendations. ## Failure modes - Orchestrator re-implementing a child skill's rules → drift and inconsistency. - Scoring an artifact that failed a hard gate as if points could redeem it. - Diagnosing a defect at the wrong failure level (e.g. restyling to fix a wrong-abstraction problem). - Shipping without a render-qa SHIP verdict. ## Validation checks - Every principle a child skill cited exists in `principles.yaml` (registry integrity). - The case file contains outputs from every skill the route required. - No SHIP verdict coexists with an unresolved hard-gate failure. - Each defect is filed at exactly one of the seven levels with an owning skill. ## Repair strategies - Route each defect to the skill that owns its level and re-run `visual-render-qa`. - Iterate gate → repair → re-render until SHIP or an ESCALATE decision is returned to the user. ## Output contract The case file + a verdict. On SHIP: the artifact plus a short justification (what works, which principles were honored). On REPAIR: the ranked defects with owning skills. On ESCALATE: the specific decision needing the user. ## Source basis Orchestration + gating: QA-01/02/03 (S10). Reasoning discipline and critique: CRIT-01..06 (S01, S10, S12). Rubric weighting per the project brief §11; failure levels per §9. Child-skill principles live in their own SKILL.md Source basis and in principles.yaml. ## Known open questions - Calibrating rubric weights against the `visualize/` validation set (later track). - Whether critique-mode should output a machine-readable defect list for tooling.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · Lmandrake/MandrakeWorkflows | 100 | MIT |
| electricsheephq/worldos | 87 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 4fbaf0980947 (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
- 100fixed-view-visual-benchmark
Renderuje wizualizację w stałym widoku i ocenia ją niezależnie według rubryki, gdy potrzebny jest powtarzalny benchmark. - 95fixed-view-visual-benchmark
Renderuje wizualizacje w stałym widoku i ocenia je niezależnie według rubryki, bez swobodnej krytyki wizualnej. - 100fixed-view-visual-benchmark
Renderuje wizualizację w stałym widoku i ocenia ją niezależnie według rubryki, gdy potrzebny jest powtarzalny benchmark wizualny. - 100design-critic
Ocenia estetykę i punktuje projekt w 6 wymiarach, do krytyki UI, przeglądu projektu, oceny jakości wizualnej i sugestii remiksu. - 100visual-asset-critic
Porównuje wygenerowany obraz z ujęciem ze storyboardu i promptem, a następnie przygotowuje uwagi do poprawek, zanim grafika zostanie zatwierdzona. - 100web-design-orchestrator
Koordynuje projekt, przeprojektowanie i wdrożenie strony lub interfejsu od odkrycia i referencji przez kierunek artystyczny po build, krytykę, QA i bramki wydania.