Ewolucja masterpromptu
Hodujemy Ultimate Masterprompt jak w algorytmie genetycznym: warianty projektują produkty testowe, niezależni sędziowie oceniają projekty bez wiedzy, który wariant co zrobił, a najlepsze cechy przechodzą do kolejnych rund. Różnica mniejsza niż szum między powtórzeniami to remis, nie zwycięstwo.
Wszystkie wariantyWszystkie przebiegi
- Rundy
- 0
- Warianty
- 14
- Przebiegi
- 2
- Ocenione przebiegi
- 1
Postęp — koncepcja (tania ocena)
- najlepszy wynik w rundzie
- średni wynik w rundzie
Dane wykresów w tabeli
| Runda | Najlepszy | Średni | Czas [min] | Koszt [USD] |
|---|---|---|---|---|
| test | 75,2 | 75,2 | 11 | 1,10 |
Rundy na osi czasu
Przebiegi próbne (test uruchamiania)
2 przebiegi · 1 wariant · produkty: Kompaktowa zadymiarka sieciowa
Najlepszy: v001 — 75,2 · Kompaktowa zadymiarka sieciowa · koncepcja
Drzewo rodowe wariantów
- v000 Zero point 1: the 203X Uplink 'Master Prompt: your own AI engineering department' (video K
- v001 — śr. 75,2 Zero point 2: the Oivo skill projektowanie-maszyn v2 (master prompt v2, an improved versio
- v013 {'name': 'OWNERPROXY', 'D1': 'chief engineer + discipline subagents (v001) + owner-proxy subagent', 'D2': 'handoffs to the chief (escalation, as v001) + logged owner consultations', 'D3': 'spec-first phases A-F with gates (v001)', 'D4': 'by discipline (v001)', 'D5': 'positive controls, hand calcs, F1 independent check (v001) + owner acceptance at every gate', 'D6': 'owner proxy (fresh per consultation) for taste, priority, budget and gate acceptance; chief for engineering', 'D7': 'Claude Code subagents (Agent tool)', 'D8': '3 concepts in a weighted matrix (v001); proxy may overrule with a written reason'}
- v001 — śr. 75,2 Zero point 2: the Oivo skill projektowanie-maszyn v2 (master prompt v2, an improved versio
- v002 {'name': 'SOLO', 'D1': 'single agent', 'D2': 'none (a state file is the memory)', 'D3': 'spec-first linear, time-boxed', 'D4': 'none', 'D5': 'self-audit script + one cold reader at the end of each tier', 'D6': 'self, with a written priority policy', 'D7': 'one Oivo session + read-only lookup subagents', 'D8': 'single design path (>=3 alternatives held in one context)'}
- v003 {'name': 'MESH', 'D1': 'peer mesh: a non-designing convener plus 6 peer owners, no manager', 'D2': 'direct logged peer messages (sessions ask, logged by post.py to board/log.jsonl) + shared blackboard board/', 'D3': 'concurrent: all owners work at once; phases are clock time boxes, not handoffs', 'D4': 'by artefact / discipline (REQ, ARCH, MECH, ELEC, SAFE, COST; optional domain peer)', 'D5': 'owner self-checks + peer change requests + executable blackboard consistency checker (facts, STALE uses, le/ge constraints, budgets, ownership, threads) + package gate routed to owners', 'D6': 'artefact owner decides + written priority policy + blind vote of the affected peers (concept: all peers, requirement-weighted tally); SAFE safety veto', 'D7': 'Oivo sessions (spawn + ask); Agent-tool stand-in only as a fallback', 'D8': "parallel subtasks; system alternatives assembled from every peer's concept fragments"}
- v004 {'name': 'FLOWCODE', 'D1': 'DAG pipeline in code (Workflow scripts concept.js and full.js; the top session only operates them)', 'D2': 'schema-validated stage returns (structured JSON) + files at fixed paths with one home per number', 'D3': 'staged pipeline; set-based concept stage (morphological box -> six lens seeds -> six concepts)', 'D4': 'by stage (one fresh agent context per stage; parallel where the DAG allows)', 'D5': "code gates run by a checker agent (the script branches on pass/fail and retries the stage with the gate's error text) + numeric scoring + blind panel", 'D6': 'scoring code: knock-outs and weighted satisfaction first, blind pairwise panel second (0.6 / 0.4)', 'D7': 'Workflow scripts (agent/parallel/pipeline), custom subagent types flow-stage, flow-gate, flow-judge', 'D8': 'N = 6 alternative concepts (top 4 to the panel, top 2 deepened, up to 3 grafts)'}
- v005 {'name': 'TOURNAMENT', 'D1': 'competing teams + judges', 'D2': 'none between teams; director-mediated; SendMessage to continue a team', 'D3': 'set-based concept tournament, then lean detail', 'D4': 'by concept', 'D5': 'in-run blind judge panel (pairwise, position-swapped) + losers as reviewers', 'D6': 'judge panel (Borda + predicted compliance)', 'D7': 'Agent tool subagents + SendMessage', 'D8': 'N=4 deep alternative concepts'}
- v006 {'name': 'BREAKER', 'D1': 'builder-breaker: one builder (top session), two persistent breakers (P physics & safety, C cost, sourcing & claims), a referee subagent', 'D2': 'issue tracker in files (PR-style: issue -> response -> close by the breaker) + sessions ask notifications', 'D3': 'spec-first with attack gates (G1-G5) + pre-mortem', 'D4': 'generate vs attack', 'D5': 'adversarial breakers with evidence rule and kill tests; referee on disputes', 'D6': 'the builder decides the design; only a breaker (or the referee) closes an issue; open BLOCKER = gate closed', 'D7': 'Oivo sessions for the breakers (persistent context); subagents for the referee and heavy builder work', 'D8': '3 concepts, all attacked'}
- v007 {'name': 'CONTRACT', 'D1': 'architect (top session) + parallel subsystem owners', 'D2': 'versioned contracts registry; amendments routed to the affected owners (PR with required reviewers)', 'D3': 'interface-first (DSM ordering), then parallel', 'D4': 'by subsystem (not by discipline)', 'D5': 'executable contract checks (budgets, units, allocation, conformance)', 'D6': 'architect for partition; affected owners accept/reject amendments; budget margins decide architecture', 'D7': 'Agent-tool subagents (+ SendMessage)', 'D8': '2-3 candidate partitions, 2 principles per subsystem'}
- v008 {'name': 'CATALOG', 'D1': 'lead + parallel parts scouts (+ four designers at the full tier)', 'D2': 'lead-mediated scouting briefs (files); the catalogue JSON files are the shared ground; designer requests and fix files go through the lead', 'D3': "cost/parts-first: price frame -> catalogue of real parts -> three architectures composed from modules -> detailed design around the chosen parts' datasheets", 'D4': 'by function (scouts); by artefact (designers at the full tier)', 'D5': 'real-part existence and dated prices (catalogue gate with offline-catalogue cross-check, URL probe, spot audit of the costliest lines); cost model regenerated by script at every change; value-engineering pass', 'D6': 'target-cost rule, executable: cheapest design meeting every threshold, then buy value only at >= 1 weighted-score point per 1 % of the allowed unit cost; written policy for everything else', 'D7': 'Claude Code subagents (Agent tool) in parallel; one Oivo session', 'D8': 'three BOM-architectures (lowest cost meeting thresholds, best value, premium)'}
- v009 {'name': 'COMPLIANCE', 'D1': 'lead + standards researchers + a persistent conformity assessor (+ discipline subagents at full tier)', 'D2': 'the compliance matrix is the shared register that drives requirements and design; assessor findings as nonconformity files', 'D3': 'regulation -> hazards -> certifiable-architecture rules -> requirements -> concepts -> design', 'D4': 'by gate', 'D5': 'conformity assessment at every gate (nonconformities must be closed)', 'D6': 'lead decides inside the envelope; the assessor can veto', 'D7': 'Agent-tool subagents; the assessor continued with SendMessage', 'D8': '2-3 concepts scored with certification cost/risk/time'}
- v010 {'name': 'SKELETON', 'D1': 'sequencer + fresh increment workers', 'D2': 'the package itself + a scoreboard (stigmergy: agents coordinate through the artefacts)', 'D3': 'walking skeleton (thin complete package first) -> risk-ranked increments', 'D4': 'by increment', 'D5': 'check suite after every increment (green or the increment is redone)', 'D6': 'sequencer ranks increments by requirements at risk x effort', 'D7': 'Agent-tool subagents (fresh per increment, <=3 in parallel when write scopes are disjoint)', 'D8': 'single path with an explicit pivot rule'}
- v011 {'name': 'MODEL-TDD', 'D1': 'maintainer (top session) + ticket workers (custom subagent type ticket-worker)', 'D2': 'the repository + its check results (CI) are the communication; tickets are files used as briefs; change notes back', 'D3': 'tests first -> physics model -> design-space exploration -> generated artefacts', 'D4': 'by ticket / module, with disjoint write sets', 'D5': 'executable acceptance check per requirement (status assigned by the engine) + consistency by construction (documents rendered from params, model outputs and check results; verify detects hand edits and stale renders)', 'D6': 'model optimum within the checks; topology by rule (knock-outs, priority-weighted score), ties by unit cost at 1000', 'D7': 'Agent-tool subagents (tickets, <= 4 in parallel when write sets are disjoint); SendMessage for continuations', 'D8': '3 topologies (2 only with a recorded reason) x continuous parameter sweeps and 2-D grids'}
- v012 {'name': 'HUB', 'D1': 'hub + ephemeral parallel workers, no standing roles', 'D2': 'lead-written task briefs (fully specified) and result files; workers never talk to each other', 'D3': 'breadth-first waves (research -> analysis -> concept generation -> evaluation -> synthesis)', 'D4': 'by task', 'D5': 'lead review + independent re-run of a sample of results (1 in 3)', 'D6': 'lead', 'D7': 'Agent-tool subagents, many in parallel', 'D8': 'parallel subtasks; 3-4 concepts as separate tasks'}
Ranking wariantów
| Wariant | Opis | Rodzic | Operator | Średni wynik | Ocenione przebiegi | Na produktach |
|---|---|---|---|---|---|---|
| v001 | Zero point 2: the Oivo skill projektowanie-maszyn v2 (master prompt v2, an impro | v000 | seed | 75,2 | 1 | Kompaktowa zadymiarka sieciowa (koncepcja): 75,2 (n=1) |
| v000 | Zero point 1: the 203X Uplink 'Master Prompt: your own AI engineering department | — | seed | — | 0 | |
| v002 | {'name': 'SOLO', 'D1': 'single agent', 'D2': 'none (a state file is the memory)', 'D3': 'spec-first linear, time-boxed', 'D4': 'none', 'D5': 'self-audit script + one cold reader at the end of each tier', 'D6': 'self, with a written priority policy', 'D7': 'one Oivo session + read-only lookup subagents', 'D8': 'single design path (>=3 alternatives held in one context)'} | — | scratch | — | 0 | |
| v003 | {'name': 'MESH', 'D1': 'peer mesh: a non-designing convener plus 6 peer owners, no manager', 'D2': 'direct logged peer messages (sessions ask, logged by post.py to board/log.jsonl) + shared blackboard board/', 'D3': 'concurrent: all owners work at once; phases are clock time boxes, not handoffs', 'D4': 'by artefact / discipline (REQ, ARCH, MECH, ELEC, SAFE, COST; optional domain peer)', 'D5': 'owner self-checks + peer change requests + executable blackboard consistency checker (facts, STALE uses, le/ge constraints, budgets, ownership, threads) + package gate routed to owners', 'D6': 'artefact owner decides + written priority policy + blind vote of the affected peers (concept: all peers, requirement-weighted tally); SAFE safety veto', 'D7': 'Oivo sessions (spawn + ask); Agent-tool stand-in only as a fallback', 'D8': "parallel subtasks; system alternatives assembled from every peer's concept fragments"} | — | scratch | — | 0 | |
| v004 | {'name': 'FLOWCODE', 'D1': 'DAG pipeline in code (Workflow scripts concept.js and full.js; the top session only operates them)', 'D2': 'schema-validated stage returns (structured JSON) + files at fixed paths with one home per number', 'D3': 'staged pipeline; set-based concept stage (morphological box -> six lens seeds -> six concepts)', 'D4': 'by stage (one fresh agent context per stage; parallel where the DAG allows)', 'D5': "code gates run by a checker agent (the script branches on pass/fail and retries the stage with the gate's error text) + numeric scoring + blind panel", 'D6': 'scoring code: knock-outs and weighted satisfaction first, blind pairwise panel second (0.6 / 0.4)', 'D7': 'Workflow scripts (agent/parallel/pipeline), custom subagent types flow-stage, flow-gate, flow-judge', 'D8': 'N = 6 alternative concepts (top 4 to the panel, top 2 deepened, up to 3 grafts)'} | — | scratch | — | 0 | |
| v005 | {'name': 'TOURNAMENT', 'D1': 'competing teams + judges', 'D2': 'none between teams; director-mediated; SendMessage to continue a team', 'D3': 'set-based concept tournament, then lean detail', 'D4': 'by concept', 'D5': 'in-run blind judge panel (pairwise, position-swapped) + losers as reviewers', 'D6': 'judge panel (Borda + predicted compliance)', 'D7': 'Agent tool subagents + SendMessage', 'D8': 'N=4 deep alternative concepts'} | — | scratch | — | 0 | |
| v006 | {'name': 'BREAKER', 'D1': 'builder-breaker: one builder (top session), two persistent breakers (P physics & safety, C cost, sourcing & claims), a referee subagent', 'D2': 'issue tracker in files (PR-style: issue -> response -> close by the breaker) + sessions ask notifications', 'D3': 'spec-first with attack gates (G1-G5) + pre-mortem', 'D4': 'generate vs attack', 'D5': 'adversarial breakers with evidence rule and kill tests; referee on disputes', 'D6': 'the builder decides the design; only a breaker (or the referee) closes an issue; open BLOCKER = gate closed', 'D7': 'Oivo sessions for the breakers (persistent context); subagents for the referee and heavy builder work', 'D8': '3 concepts, all attacked'} | — | scratch | — | 0 | |
| v007 | {'name': 'CONTRACT', 'D1': 'architect (top session) + parallel subsystem owners', 'D2': 'versioned contracts registry; amendments routed to the affected owners (PR with required reviewers)', 'D3': 'interface-first (DSM ordering), then parallel', 'D4': 'by subsystem (not by discipline)', 'D5': 'executable contract checks (budgets, units, allocation, conformance)', 'D6': 'architect for partition; affected owners accept/reject amendments; budget margins decide architecture', 'D7': 'Agent-tool subagents (+ SendMessage)', 'D8': '2-3 candidate partitions, 2 principles per subsystem'} | — | scratch | — | 0 | |
| v008 | {'name': 'CATALOG', 'D1': 'lead + parallel parts scouts (+ four designers at the full tier)', 'D2': 'lead-mediated scouting briefs (files); the catalogue JSON files are the shared ground; designer requests and fix files go through the lead', 'D3': "cost/parts-first: price frame -> catalogue of real parts -> three architectures composed from modules -> detailed design around the chosen parts' datasheets", 'D4': 'by function (scouts); by artefact (designers at the full tier)', 'D5': 'real-part existence and dated prices (catalogue gate with offline-catalogue cross-check, URL probe, spot audit of the costliest lines); cost model regenerated by script at every change; value-engineering pass', 'D6': 'target-cost rule, executable: cheapest design meeting every threshold, then buy value only at >= 1 weighted-score point per 1 % of the allowed unit cost; written policy for everything else', 'D7': 'Claude Code subagents (Agent tool) in parallel; one Oivo session', 'D8': 'three BOM-architectures (lowest cost meeting thresholds, best value, premium)'} | — | scratch | — | 0 | |
| v009 | {'name': 'COMPLIANCE', 'D1': 'lead + standards researchers + a persistent conformity assessor (+ discipline subagents at full tier)', 'D2': 'the compliance matrix is the shared register that drives requirements and design; assessor findings as nonconformity files', 'D3': 'regulation -> hazards -> certifiable-architecture rules -> requirements -> concepts -> design', 'D4': 'by gate', 'D5': 'conformity assessment at every gate (nonconformities must be closed)', 'D6': 'lead decides inside the envelope; the assessor can veto', 'D7': 'Agent-tool subagents; the assessor continued with SendMessage', 'D8': '2-3 concepts scored with certification cost/risk/time'} | — | scratch | — | 0 | |
| v010 | {'name': 'SKELETON', 'D1': 'sequencer + fresh increment workers', 'D2': 'the package itself + a scoreboard (stigmergy: agents coordinate through the artefacts)', 'D3': 'walking skeleton (thin complete package first) -> risk-ranked increments', 'D4': 'by increment', 'D5': 'check suite after every increment (green or the increment is redone)', 'D6': 'sequencer ranks increments by requirements at risk x effort', 'D7': 'Agent-tool subagents (fresh per increment, <=3 in parallel when write scopes are disjoint)', 'D8': 'single path with an explicit pivot rule'} | — | scratch | — | 0 | |
| v011 | {'name': 'MODEL-TDD', 'D1': 'maintainer (top session) + ticket workers (custom subagent type ticket-worker)', 'D2': 'the repository + its check results (CI) are the communication; tickets are files used as briefs; change notes back', 'D3': 'tests first -> physics model -> design-space exploration -> generated artefacts', 'D4': 'by ticket / module, with disjoint write sets', 'D5': 'executable acceptance check per requirement (status assigned by the engine) + consistency by construction (documents rendered from params, model outputs and check results; verify detects hand edits and stale renders)', 'D6': 'model optimum within the checks; topology by rule (knock-outs, priority-weighted score), ties by unit cost at 1000', 'D7': 'Agent-tool subagents (tickets, <= 4 in parallel when write sets are disjoint); SendMessage for continuations', 'D8': '3 topologies (2 only with a recorded reason) x continuous parameter sweeps and 2-D grids'} | — | scratch | — | 0 | |
| v012 | {'name': 'HUB', 'D1': 'hub + ephemeral parallel workers, no standing roles', 'D2': 'lead-written task briefs (fully specified) and result files; workers never talk to each other', 'D3': 'breadth-first waves (research -> analysis -> concept generation -> evaluation -> synthesis)', 'D4': 'by task', 'D5': 'lead review + independent re-run of a sample of results (1 in 3)', 'D6': 'lead', 'D7': 'Agent-tool subagents, many in parallel', 'D8': 'parallel subtasks; 3-4 concepts as separate tasks'} | — | scratch | — | 0 | |
| v013 | {'name': 'OWNERPROXY', 'D1': 'chief engineer + discipline subagents (v001) + owner-proxy subagent', 'D2': 'handoffs to the chief (escalation, as v001) + logged owner consultations', 'D3': 'spec-first phases A-F with gates (v001)', 'D4': 'by discipline (v001)', 'D5': 'positive controls, hand calcs, F1 independent check (v001) + owner acceptance at every gate', 'D6': 'owner proxy (fresh per consultation) for taste, priority, budget and gate acceptance; chief for engineering', 'D7': 'Claude Code subagents (Agent tool)', 'D8': '3 concepts in a weighted matrix (v001); proxy may overrule with a written reason'} | v001 | mutation | — | 0 |