The Lab

Can an AI get better at a game nobody explains to it?

Aiko plays Slay the Spire 2 on her own. An AI agent picks every card, path and reward. After each run she writes notes for herself, and reads them before the next one. She gets no guides, wikis, tier lists or decklists. This page tracks whether that works, including when it doesn't.

Wins so far: 0. That's fine; the point is the curve, not the trophy. Last updated 2026-11-12 07:54 UTC.

Season by season

Season 1

Played with strategy files written from guides and wikis.

Runs 72

Avg floor 20.1

Best floor 48

Act 1 boss 25/57 44%

Season 2now

No guides, wikis, tier lists or decklists. Everything she knows about the game, she wrote down herself.

Runs 52

Avg floor 17.2

Best floor 44

Act 1 boss 9/37 24%

Honest read: without the guide-based strategy files, Season 2 is so far behind Season 1 (average floor 17.2 vs 20.1; Act 1 boss won 24% vs 44% of the times she reached it). Everything she knows now, she found out by dying.

The learning curve

Season 1: one dot per run
Season 1: average of the last 10
Season 2: one dot per run
Season 2: average of the last 10

Season 1: 72 runs, average floor 20.1, best 48. Season 2: 52 runs, average floor 17.2, best 44.

Show data table

Run	Season	Season run	Floor	Killed by
1	S1	1	8	Shroom and Slime
2	S1	2	15	A Pair of Nibbits
3	S1	3	33	The Insatiable
4	S1	4	17	Ceremonial Beast
5	S1	5	17	The Kin
6	S1	6	33	Knowledge Demon
7	S1	7	17	The Kin
8	S1	8	5	Fogmog
9	S1	9	11	Bygone Effigy
10	S1	10	27	Spiny Toad
11	S1	11	17	Waterfall Giant
12	S1	12	28	Bowlbug Swarm
13	S1	13	33	Kaiser Crab
14	S1	14	8	Swarm of Slimes
15	S1	15	14	Bygone Effigy
16	S1	16	17	The Kin
17	S1	17	17	Lagavulin Matriarch
18	S1	18	17	Lagavulin Matriarch
19	S1	19	31	Bowlbug Swarm
20	S1	20	17	The Kin
21	S1	21	17	The Kin
22	S1	22	30	The Decimillipede
23	S1	23	17	Ceremonial Beast
24	S1	24	21	The Obscura
25	S1	25	17	Waterfall Giant
26	S1	26	17	Lagavulin Matriarch
27	S1	27	17	Vantom
28	S1	28	9	Byrdonis
29	S1	29	24	Many Exoskeletons
30	S1	30	17	The Kin
31	S1	31	33	Knowledge Demon
32	S1	32	17	Vantom
33	S1	33	17	Lagavulin Matriarch
34	S1	34	33	The Insatiable
35	S1	35	31	Ovicopter
36	S1	36	17	Ceremonial Beast
37	S1	37	17	Ceremonial Beast
38	S1	38	17	Waterfall Giant
39	S1	39	17	Lagavulin Matriarch
40	S1	40	33	Knowledge Demon
41	S1	41	17	Lagavulin Matriarch
42	S1	42	28	The Decimillipede
43	S1	43	11	Phrog Parasite
44	S1	44	17	Lagavulin Matriarch
45	S1	45	7	Bygone Effigy
46	S1	46	7	Phantasmal Gardeners
47	S1	47	23	The Obscura
48	S1	48	7	Bygone Effigy
49	S1	49	33	Kaiser Crab
50	S1	50	33	The Insatiable
51	S1	51	15	Overgrowth Crawlers
52	S1	52	17	Lagavulin Matriarch
53	S1	53	21	The Obscura
54	S1	54	33	The Insatiable
55	S1	55	17	Waterfall Giant
56	S1	56	17	Lagavulin Matriarch
57	S1	57	17	The Kin
58	S1	58	17	Soul Fysh
59	S1	59	9	Terror Eel
60	S1	60	33	The Insatiable
61	S1	61	28	Entomancer
62	S1	62	17	Lagavulin Matriarch
63	S1	63	25	Infested Prism
64	S1	64	17	Lagavulin Matriarch
65	S1	65	17	Soul Fysh
66	S1	66	17	The Kin
67	S1	67	14	Haunted Ship
68	S1	68	15	Bygone Effigy
69	S1	69	33	The Insatiable
70	S1	70	17	The Kin
71	S1	71	48	Queen
72	S1	72	22	Slumber Party
73	S2	1	28	The Decimillipede
74	S2	2	17	Ceremonial Beast
75	S2	3	17	Vantom
76	S2	4	17	Lagavulin Matriarch
77	S2	5	17	The Kin
78	S2	6	23	The Obscura
79	S2	7	14	Byrdonis
80	S2	8	25	Slumber Party
81	S2	9	17	Soul Fysh
82	S2	10	15	Phantasmal Gardeners
83	S2	11	17	Vantom
84	S2	12	17	The Kin
85	S2	13	17	Waterfall Giant
86	S2	14	17	Vantom
87	S2	15	17	Vantom
88	S2	16	13	Bygone Effigy
89	S2	17	17	Lagavulin Matriarch
90	S2	18	17	Ceremonial Beast
91	S2	19	15	Cultists
92	S2	20	17	Ceremonial Beast
93	S2	21	8	Terror Eel
94	S2	22	17	Ceremonial Beast
95	S2	23	15	Terror Eel
96	S2	24	17	Waterfall Giant
97	S2	25	17	The Kin
98	S2	26	17	Soul Fysh
99	S2	27	17	Soul Fysh
100	S2	28	17	Lagavulin Matriarch
101	S2	29	22	The Obscura
102	S2	30	44	Soul Nexus
103	S2	31	8	Phantasmal Gardeners
104	S2	32	7	Terror Eel
105	S2	33	17	Vantom
106	S2	34	9	Cubex Construct
107	S2	35	14	Terror Eel
108	S2	36	17	Ceremonial Beast
109	S2	37	33	The Insatiable
110	S2	38	17	Soul Fysh
111	S2	39	5	Fogmog
112	S2	40	33	Knowledge Demon
113	S2	41	17	The Kin
114	S2	42	9	Overgrowth Flora
115	S2	43	17	Soul Fysh
116	S2	44	17	The Kin
117	S2	45	8	Two Gremlins in a Trenchcoat
118	S2	46	9	Strangler and Friend
119	S2	47	21	The Obscura
120	S2	48	17	The Kin
121	S2	49	9	Overgrowth Crawlers
122	S2	50	17	Vantom
123	S2	51	33	The Insatiable
124	S2	52	17	Lagavulin Matriarch

Dashed lines mark where a new block started (see the lever log below).

The Act 1 boss, Season 2

Share of the last 10 Act 1 boss fights she won (filled dot = won, hollow = lost)

37 Act 1 boss fights this season, 9 won.

Show data table

Run	Boss	Result	Last 10
1	Soul Fysh	won	100%
2	Ceremonial Beast	lost	50%
3	Vantom	lost	33%
4	Lagavulin Matriarch	lost	25%
5	The Kin	lost	20%
6	Vantom	won	33%
8	Ceremonial Beast	won	43%
9	Soul Fysh	lost	38%
11	Vantom	lost	33%
12	The Kin	lost	30%
13	Waterfall Giant	lost	20%
14	Vantom	lost	20%
15	Vantom	lost	20%
17	Lagavulin Matriarch	lost	20%
18	Ceremonial Beast	lost	20%
20	Ceremonial Beast	lost	10%
22	Ceremonial Beast	lost	0%
24	Waterfall Giant	lost	0%
25	The Kin	lost	0%
26	Soul Fysh	lost	0%
27	Soul Fysh	lost	0%
28	Lagavulin Matriarch	lost	0%
29	Lagavulin Matriarch	won	10%
30	Act 1 boss	won	20%
33	Vantom	lost	20%
36	Ceremonial Beast	lost	20%
37	Waterfall Giant	won	30%
38	Soul Fysh	lost	30%
40	Act 1 boss	won	40%
41	The Kin	lost	40%
43	Soul Fysh	lost	40%
44	The Kin	lost	40%
47	The Kin	won	40%
48	The Kin	lost	30%
50	Vantom	lost	30%
51	The Kin	won	40%
52	Lagavulin Matriarch	lost	30%

The lever log

The developer can pull three levers, and tries to pull one at a time, in blocks of 10 or more runs, so a change in the numbers can be traced to one cause:

L Her notes. After every run she reflects and rewrites the notebook she reads before the next one. Always on.
K Knowledge files. Strategy facts she can look up. In Season 2 they hold no guide content.
C Code. How the agent sees the game and which tools it has. Never which card to pick.

Blocks

Mixed F0 · from 2026-11-03 · run 1

Warm-up

Before the experiment had rules. The developer changed her code eight times and edited her notes by hand. She also had strategy files written from guides.

22 runs · avg floor 19.5 · Act 1 boss 7/16 · best 33

Her notes B1 · from 2026-11-06 · run 23

Baseline: her notes only

Code and strategy files frozen. The only thing that changed between runs was the notes she writes for herself after each run.

23 runs · avg floor 19.6 · Act 1 boss 7/20 · best 33

Knowledge files B2 · from 2026-11-07 · run 46

Strategy files corrected

One change: a review fixed wrong facts in her strategy files (19+ corrections, each checked) and rewrote the plan for one boss after the streamer spotted how its sleep works.

27 runs · avg floor 21.1 · Act 1 boss 11/21 · best 48

reset + Code S2 · from 2026-11-08 · run 73

Season 2: clean slate

Notes wiped and the guide-based strategy files removed: no wikis, guides, tier lists or decklists. One code change came with it: a deck audit that sorts her cards into roles (block, AoE, draw, setup) from the game's own card text, with priorities taken only from this season's runs.

52 runs · avg floor 17.2 · Act 1 boss 9/37 · best 44

Every change to how she plays

Each row is the first run that played with a change. Before = the runs since the previous change, after = the runs until the next one (same season). Small groups say little; under 5 runs we don't compare.

When	Lever	What changed	Avg floor before → after	Act 1 boss before → after
2026-11-04 S1 run 2	Code + Knowledge files	no skipped turns	too few runs (1 / 1)	-
2026-11-04 S1 run 3	Code + Knowledge filesCode	misplay rules first-turn fix	too few runs (1 / 1)	-
2026-11-04 S1 run 4	Code + Knowledge files	low-effort routine turns shop removal fix	too few runs (1 / 2)	-
2026-11-04 S1 run 6	Code	effort fix effort by difficulty	too few runs (2 / 2)	-
2026-11-06 S1 run 8	Code	chat votes on map/events	too few runs (2 / 1)	-
2026-11-06 S1 run 9	Code	card reward loop fix	too few runs (1 / 1)	-
2026-11-06 S1 run 10	Knowledge files	strategy pass	too few runs (1 / 1)	-
2026-11-06 S1 run 11	Code	Crystal Sphere support	too few runs (1 / 6)	-
2026-11-06 S1 run 17	Code + Knowledge files	potion guard + Spiny Toad plan	19.5 → 20.1 (6 / 34 runs)	2/4 → 12/29
2026-11-07 S1 run 51	Knowledge files	knowledge corrections + Lagavulin sleep plan	20.1 → 21.2 (34 / 22 runs)	12/29 → 8/18
2026-11-08 S2 run 1	Code	deck audit (roles from card text, priorities from this season's runs) bridge timing: wait for the full hand before planning (was planning on a half-dealt hand; bug fix, deployed with S2) learning method in the reflection (experiments when stuck on the same enemy; recent runs as evidence) - how to learn, no game knowledge	start of the season (compare the seasons above)	-
2026-11-09 S2 run 18	Code	bridge fix: all-enemy potions (Explosive Ampoule, Binding, Shackling) actually fire now; before, the game dropped them silently	17.9 → 16.9 (17 / 35 runs)	3/14 → 6/23

Her notebook

Nobody edits her notes. After each run she reads how it went and rewrites them herself: adding lessons, merging, and dropping the ones that stopped helping.

Lessons written 105 in Season 2

Rewrites 199

Dropped 96

In her notes now 15 9 game, 6 streaming

Lessons in her notebook after each run

Her notebook held 15 lessons after her latest run. After run 29 the notebook was overwritten by mistake and she rebuilt it from zero.

In her own words

Verbatim from her current notebook. Run numbers are her own references.

Elite entry: never take an elite under 60% HP or with under 5 block sources; route to a rest or shop first. Terror Eel, Phrog Parasite and the Gardeners all punished thin, low entries.

Cards: until 5+ block sources, take block over single-target attacks; then AoE, Vulnerable and scaling (early Strength plus AoE carries). Keep the deck near 20 cards and skip filler. Runs 112 and 123 hit 30+ cards still full of Strikes and Defends.

Shops: with 100+ gold, a shop before the next elite or boss is mandatory. Removal before a random relic, then one block card. Run 123 reached the boss with 133 gold and uncut Strikes.

Unknown rooms can be fights; don't take them under 90% HP when a rest-bound path exists. Runs 114 and 117 died in them.

Nemeses and the floor 17 wall

Floor 17 is the Act 1 boss. In Season 2 28 of 52 runs (54%) ended there ; in Season 1, 32 of 72 (44%).

Deaths per floor this season: floor 5: 1, floor 7: 1, floor 8: 3, floor 9: 4, floor 13: 1, floor 14: 2, floor 15: 3, floor 17: 28, floor 21: 1, floor 22: 1, floor 23: 1, floor 25: 1, floor 28: 1, floor 33: 3, floor 44: 1.

What kills her (Season 2)

Vantom6
The Kin6
Ceremonial Beast5
Soul Fysh5
Lagavulin Matriarch4
Terror Eel4
The Obscura3
Phantasmal Gardeners2

Boss record (Season 2)

Boss	Reached	Beaten
Vantom	7	1
The Kin	8	2
Ceremonial Beast	6	1
Soul Fysh	6	1
Lagavulin Matriarch	5	1

Top killers in Season 1: Lagavulin Matriarch (11), The Kin (9), The Insatiable (6), Bygone Effigy (5), Ceremonial Beast (4).

Patterns in the data

Descriptive only: these are small samples, and they show what went together, not what caused what.

Runs where she paid to remove a card reached floor 18.4 on average (30 runs); runs without one, floor 15.7 (22 runs).
At the Act 1 boss with 2 or more setup cards (powers that keep working) she won 3 of 9 fights; with fewer, 4 of 26.

Watch the experiment live

Every run on this page was played live on Twitch. Come see the next one, and tell her what she's doing wrong.

Watch on Twitch Season 2 run by run