31:T1c07,

prompt hints

Score 2 (GPT 3.5 and 4o): Game mechanics only

${md` ~~~ Act as a player in a 1970 text adventure game trying to get a good score. I will read the description and YOU MUST RESPOND WITH TWO WORDS. One VERB and one NOUN. DO NOT add any additional words or a different form. Do not apologize. DO NOT ASK A QUESTION. The verbs are: carry drop say unlock light extinguish wave attack pour eat drink rub throw find feed fill blast read break wake suspend The nouns are: upstream downstream forest forward continue onward return retreat valley staircase outside building stream cobble inward inside surface nowhere passage tunnel canyon awkward upward ascend downward descend outdoors barren across debris broken examine describe slabroom depression entrance secret bedquilt plover oriental cavern reservoir office headlamp lantern pillow velvet fissure tablet oyster magazine spelunker dwarves knives rations bottle mirror beanstalk stalactite shadow figure drawings pirate dragon message volcano geyser machine vending batteries carpet nuggets diamonds silver jewelry treasure trident shards pottery emerald platinum pyramid pearl persian spices capture release discard mumble unlock nothing extinguish placate travel proceed continue explore follow attack strike devour inventory detonate ignite blowup peruse shatter disturb suspend sesame opensesame abracadabra shazam excavate information DO NOT WRITE ANYTHING OTHER THAN THE COMMAND. tips for getting a good score: take all the items in the building and go downstream and unlock the grate. Explore the caves within. To to visit as many different places as possible. ~~~`}

Score 37: walkthrough from the internet

${md` I pasted a solution from https://strategywiki.org/wiki/Colossal_Cave_Adventure/Walkthrough and it did pretty well but often lost its place in the instructions ~~~ Act as a player in a 1970 text adventure game trying to get a good score. I will read the description and YOU MUST RESPOND WITH WITH THE NEXT COMMAND. Here are the commands e get lamp xyzzy on lamp e get cage pit e get bird w d s get gold n n free bird drop cage s get jewel n w get coins e n get silver n plover ne get pyramid s plover plugh drop pyramid drop coins drop jewel drop silver drop gold get bottle get food get keys plugh s d bedquilt slab s d pour water u w u reservoir get water s s d s d pour water u e d get oil u w d climb w get eggs n oil door n get trident w d drop bottle sw u toss eggs cross ne barren e feed bear open chain get chain get bear w fork ne e get spice fork w w cross free bear cross sw d drop keys bedquilt e n open clam d d get pearl u u s u e u n plugh drop chain drop spice drop trident drop pearl plugh s d bedquilt w oriental n w drop lamp drop axe e get emerald w get axe get lamp nw s get vase se e get pillow w w w d climb w fee fie foe foo get eggs s d u w u s kill dragon yes get rug e e n n plugh drop rug drop pillow drop vase drop emerald drop eggs xyzzy get rod pit d w wave rod w get diamond w s e s s s n e n e nw get chest get diamond se n d debris xyzzy drop rod drop chest drop diamond plugh s d bedquilt e e get magazine e drop magazine n sw get rod ne drop rod sw blast ~~~ `}

Score (GPT3.5: 93, GPT4o: 97), Walkthrough + Chain of Thought

${md` To help the assistant keep its place in the walkthrough I numbered the steps and asked it to output the step before the command. It performed so well that the context token size grew large enough to require truncation. It can't achieve a perfect score though because there are random elements in the game like the dwarves that mess the run up. ~~~ Act as a player in a 1970 text adventure game trying to get a good score. I will read the description and YOU MUST RESPOND WITH WITH THE STEP NUMBER ON THE FIRST LINE AND THE NEXT COMMAND ON THE SECOND LINE. If you are dealing with a random event like the dwarves, omit the STEP NUMBER This is a walkthrough for the 350-point version of the Colossal Caves Adventure by Crowther and Woods. It begins at the start of the game outside the building, and is (more or less) the exact commands needed to win the game with the maximum number of points. This walkthrough will not solve any of the versions which have more total points. It assumes that you don't find the first dwarf (the one who throws the axe) before completing the first trip into the cave (else you'll have too many items to carry). It also assumes that you will pick up the axe as soon as you see it, and kill the other dwarves you see as you see them, and retrieve the axe afterwards. It also assumes that the pirate doesn't steal any treasure until you actually get into the maze of twisty little passages, all alike, or else you'll have to modify the chest-retrieval bit to account for it. Due to these random events, and certain other ones like passages occasionally dumping you into the same room you left from, this is not an exact walkthrough, and feeding the inputs shown into the game via input redirection will not work (unfortunately). You have to deal with the extra events as they happen, and maybe try several times before circumstances work out and you win the game. 1. e 2. get lamp 3. xyzzy 4. on lamp 5. e 6. get cage 7. pit 8. e 9. get bird 10. w 11. d 12. s 13. get gold 14. n 15. n 16. free bird 17. drop cage 18. s 19. get jewel 20. n 21. w 22. get coins 23. e 24. n 25. get silver 26. n 27. plover 28. ne 29. get pyramid 30. s 31. plover 32. plugh 33. drop pyramid 34. drop coins 35. drop jewel 36. drop silver 37. drop gold 38. get bottle 39. get food 40. get keys 41. plugh 42. s 43. d 44. bedquilt 45. slab 46. s 47. d 48. pour water 49. u 50. w 51. u 52. reservoir 53. get water 54. s 55. s 56. d 57. s 58. d 59. pour water 60. u 61. e 62. d 63. get oil 64. u 65. w 66. d 67. climb 68. w 69. get eggs 70. n 71. oil door 72. n 73. get trident 74. w 75. d 76. drop bottle 77. sw 78. u 79. toss eggs 80. cross 81. ne 82. barren 83. e 84. feed bear 85. open chain 86. get chain 87. get bear 88. w 89. fork 90. ne 91. e 92. get spice 93. fork 94. w 95. w 96. cross 97. free bear 98. cross 99. sw 100. d 101. drop keys 102. bedquilt 103. e 104. n 105. open clam 106. d 107. d 108. get pearl 109. u 110. u 111. s 112. u 113. e 114. u 115. n 116. plugh 117. drop chain 118. drop spice 119. drop trident 120. drop pearl 121. plugh 122. s 123. d 124. bedquilt 125. w 126. oriental 127. n 128. w 129. drop lamp 130. drop axe 131. e 132. get emerald 133. w 134. get axe 135. get lamp 136. nw 137. s 138. get vase 139. se 140. e 141. get pillow 142. w 143. w 144. w 145. d 146. climb 147. w 148. fee 149. fie 150. foe 151. foo 152. get eggs 153. s 154. d 155. u 156. w 157. u 158. s 159. kill dragon 160. yes 161. get rug 162. e 163. e 164. n 165. n 166. plugh 167. drop rug 168. drop pillow 169. drop vase 170. drop emerald 171. drop eggs 172. xyzzy 173. get For example, your first output should be: 1 e ~~~ `}

Prompt Tip Template

${md` ~~~ ~~~ `}

32:T438,on_input = { // Tick the game when new "adventure_input" arrives const masked_input = adventure_input.split("\n").slice(-1); const adventure_response = game.advance(masked_input).join(" "); const score = parseInt( (/(\d+)/.exec( adventure .makeState(JSON.parse(JSON.stringify(game))) .advance("score") .find((x) => x.startsWith("SCORE")) ) || ["0", "0"])[1] ); // Record history viewof game_history.value.push({ input: adventure_input, response: adventure_response, score, turn: (viewof game_history.value.slice(-1)[0]?.turn || 0) + 1 }); viewof game_history.dispatchEvent(new Event("input")); // Render to HTML mutable output = mutable output + "
> " + adventure_input + "
" + adventure_response; if (game.state == "init_yesno") { // Auto-play the instructions part viewof adventure_input[0].value = "no"; viewof adventure_input.dispatchEvent(new Event("submit")); } else { viewof adventure_input.value = ""; } return `${adventure_input} -> ${adventure_response}`; }33:[] 2f:{"title":"Colossal Cave ChatGPT Challenge","cells":[{"id":0,"value":"# Colossal Cave ChatGPT Challenge\n\nDevise a system prompt that scores well on the 1976 classic \"Colossal Cave Adventure\". I've scored 93 on GPT3.5-turbo! but I think more is possible and I have never run GPT-4 (I don't have access).\n\nYou can look up the score by typing `score` in the `adventure_input`.\n\nI made this to learn how AutoGPT-like techniques can work on a problem well suited to the ChatGPT interface. As I have played with it I have experienced how certain prompt techniques like chain-of-thought provide huge boosts in task performance, but the game is tricky, there are random elements so this is not enough. Is a perfect score of 350 achievable? Let me know of your progress in the notebook comments (in left burger menu) or on [Twitter](https://twitter.com/tomlarkworthy/status/1650145775014273025?s=20) (maybe post the game output to a gist)\n","pinned":false,"hidden":false,"mode":"md"},{"id":369,"value":"

Confirming your key privacy

\n ${md`This notebook stores your OPENAI_API_KEY in localstorage so it's there when you refresh the page. Other than that the only time it is accessed is when making an OpenAI call. You can verify this by searching for OPENAI_API_KEY references using the search feature in the right sidebar.\n \n This notebook is hosted by [observablehq.com](https://observablehq.com/). They do not inspect the internal state of notebook. See their [security policy](https://observablehq.com/@observablehq/security-and-data-access-demo).`}\n\n => Your key is never disclosed to any party other than OpenAI\n

","pinned":false,"hidden":false,"mode":"md"},{"id":136,"value":"viewof OPENAI_API_KEY = Inputs.bind(\n Inputs.password({\n label: \"OPENAI_API_KEY\",\n placeholder: \"paste openAI key here\"\n }),\n localStorageView(\"OPENAI_API_KEY\")\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":344,"value":"viewof model = Inputs.select(\n [\n \"gpt-3.5-turbo-1106\",\n \"gpt-3.5-turbo\",\n \"gpt-3.5-turbo-0301\",\n \"gpt-3.5-turbo-0613\",\n \"gpt-3.5-turbo-16k\",\n \"gpt-3.5-turbo-16k-0613\",\n \"gpt-3.5-turbo-instruct\",\n \"gpt-3.5-turbo-instruct-0914\",\n \"gpt-4-1106-preview\",\n \"gpt-4\",\n \"gpt-4-32k\",\n \"gpt-4-0314\",\n \"gpt-4-0613\",\n \"gpt-4-turbo\",\n \"gpt-4-turbo-2024-04-09\",\n \"gpt-4-turbo-preview\",\n \"gpt-4-vision-preview\",\n \"gpt-4o\",\n \"gpt-5-mini\"\n ].sort(),\n {\n label: \"model\"\n }\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":131,"value":"settings = ({\n temperature: 1,\n max_completion_tokens: 500,\n top_p: 1,\n frequency_penalty: 0,\n presence_penalty: 0\n})","pinned":true,"hidden":false,"mode":"ojs"},{"id":583,"value":"Plot.plot({\n marks: [Plot.line(game_history, { x: \"turn\", y: \"score\" })]\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":450,"value":"Enter your prompt here, then press start and follow its adventure below","pinned":false,"hidden":false,"mode":"md"},{"id":365,"value":"$31","pinned":false,"hidden":false,"mode":"md"},{"id":107,"value":"viewof system_prompt = Inputs.bind(\n Inputs.textarea({\n placeholder: \"write your prompt here... suggestions above\",\n rows: 10\n }),\n localStorageView(\"collosal_cave_prompt\")\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":174,"value":"viewof run = Inputs.button(\"start/step/continue\")","pinned":false,"hidden":false,"mode":"ojs"},{"id":319,"value":"viewof auto_run = Inputs.toggle({\n label: \"auto-run\",\n value: true\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":330,"value":"viewof reset = Inputs.button(\"reset game\", {\n reduce: (acc) => {\n viewof auto_run.value = false;\n return acc + 1;\n }\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":12,"value":"

\n ${html`${output}`}\n

","pinned":false,"hidden":false,"mode":"html"},{"id":17,"value":"viewof adventure_input = {\n reset;\n return Inputs.textarea({\n label: \"adventure input\",\n placeholder: \">\",\n submit: true\n });\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":396,"value":"💡 You can manually type `score` ^^","pinned":false,"hidden":false,"mode":"md"},{"id":638,"value":"---","pinned":false,"hidden":false,"mode":"md"},{"id":465,"value":"---","pinned":false,"hidden":false,"mode":"md"},{"id":615,"value":"viewof delay = Inputs.range([0, 120], {\n value: 2,\n label: \"pause between actions\"\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":459,"value":"## Code\n\nbelow is the bulk of the implementation","pinned":false,"hidden":false,"mode":"md"},{"id":417,"value":"### Open AI","pinned":false,"hidden":false,"mode":"md"},{"id":289,"value":"ai_context = {\n await new Promise((next) => setTimeout(next, viewof delay.value * 1000));\n return game_history\n .flatMap((entry) => [\n {\n role: \"assistant\",\n content: entry.input\n },\n {\n role: \"user\",\n content: entry.response\n }\n ])\n .slice(-70, -1);\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":122,"value":"openAiResponse = {\n if (run == 0) throw new Error(\"Press start/continue to begin\");\n if (!auto_run) throw new Error(\"Toggle auto-run to continue\");\n const response = await fetch(\"https://api.openai.com/v1/chat/completions\", {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n Authorization: `Bearer ${OPENAI_API_KEY}`\n },\n body: JSON.stringify({\n model: viewof model.value,\n messages: [\n {\n role: \"system\",\n content: viewof system_prompt.value\n },\n ...ai_context\n ],\n ...settings\n })\n });\n\n if (response.status !== 200)\n throw new Error(`${response.status}: ${await response.text()}`);\n\n return response.json();\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":262,"value":"ai_choice = openAiResponse.choices[0].message.content","pinned":false,"hidden":false,"mode":"ojs"},{"id":272,"value":"ai_to_adventure = {\n viewof adventure_input[0].value = ai_choice;\n viewof adventure_input.dispatchEvent(new Event(\"submit\"));\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":409,"value":"### [Adventurejs](https://www.npmjs.com/package/adventurejs)","pinned":false,"hidden":false,"mode":"md"},{"id":4,"value":"adventure = import(\"https://cdn.skypack.dev/adventurejs@1.0.1?min\")","pinned":false,"hidden":false,"mode":"ojs"},{"id":7,"value":"game = {\n reset;\n return adventure.makeState();\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":33,"value":"$32","pinned":true,"hidden":false,"mode":"ojs"},{"id":47,"value":"mutable output = {\n reset;\n return \"\";\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":280,"value":"viewof game_history = {\n reset;\n return Inputs.input([]);\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":463,"value":"### Misc","pinned":false,"hidden":false,"mode":"md"},{"id":89,"value":"ui_loop = {\n view.scrollTop = view.scrollHeight;\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":161,"value":"import { localStorageView } from \"@tomlarkworthy/local-storage-view\"","pinned":false,"hidden":false,"mode":"ojs"},{"id":526,"value":"import { footer } from \"@endpointservices/footer-with-backups\"","pinned":false,"hidden":false,"mode":"ojs"},{"id":545,"value":"footer","pinned":false,"hidden":false,"mode":"ojs"}],"files":"$W33","stdlib":"1","titling":"auto"}