An AI prompt that deciphers hermetic texts by translating symbols and their context taking into respect the context of the symbol and vice versa.
SPOILER ALERT: Only for scientific use!!!
DANGER: Don't use the content of this repository if you don't know what you are doing.
The results are mostly unsorted! They have been created with several different versions of the prompt.
prompts/hermeneutic_circle.md is the optimized protocol. It merges the earlier prompts into one hermeneutic circle with two methods — known (symbol meanings are known) and lost (meanings are derived from structure only, then compared with tradition) — and contains the guide for the MCP server (Part III).
mcp-server/ runs the protocol over whole books. The calling model does the translation work; the server splits, stores, checks and assembles. One tool per step:
| Step | Tool | What happens |
|---|---|---|
| 1 | step1_split_corpus |
the text is split into sections that fit a 1,000,000-token context window (cut at chapter breaks) and into chunks |
| 2 | step2_translate_section |
the model reads the whole section (pre-understanding), translates it chunk by chunk into human-readable language and reviews it |
| 3 | step3_distill_section |
the model distils each section to about 100,000 tokens |
| 4 | step4_merge_distillates |
the server merges the distillates; the model writes the canon of the whole work |
| 5 | step5_retranslate_section |
the model translates every chunk anew from the source in the light of the whole work |
| 6 | step6_assemble_work |
the server assembles the complete translation and returns its file path |
Helper tools: read_artifact (paged access to sources and intermediate results), run_status (progress and the exact next call; resume after a context reset). Every step writes its intermediate results as .md files to runs/<run_id>/.
Setup:
cd mcp-server && npm installClaude Code picks the server up from .mcp.json when started in this repository. For other clients add node /absolute/path/to/mcp-server/src/index.js as a stdio MCP server. Then ask the model to decipher a text, for example: "Decipher archive/sacred texts.txt into German with the hermetics-decoder tools." The full protocol is also available as the MCP prompt hermeneutic_circle. Tests: cd mcp-server && npm test.
prompt.txt contains the original AI prompt
To decipher the complete text in a whole and not chapterwise, a 2MB context window of the AI (about 500.000 tokens) is required.
Agent mode is advised. Large books will be really expensive when using an API, they will be about 1500$-2500$. When using Visual Studio Code and Github Copilot Plus with a last-gen model (Opus 4.6 or GPT 5.4), this will be reduces to a few cents, but it won't be able to handle large books in a whole. Instead it will decipher the text chapterwise.
If you are bored and don't know hat to do, you could for instance try to change the prompt so it uses 7 iterations in an inner loop and 3 iterations in an outer loop, applying the new insights about archetypes from the inner loop to the original text in each outer iteration.
Instead you could also apply the new insights from archetypes after the last loop to the original text instead of applying it after each outer iteration, examining how it modifies it.
Modify the prompt and compare results. Instruct the AI to track iterations and modifications of archetypes per iteration for deeper insights and so on.
Another way would be ignoring the fixed loops and let them iterate until the system converges.
Try prompt_343_iterations_known_archetypes.txt and prompt_343_iterations_unknown_archetypes.txt for a 7³ steps system
You could also try prompt_magnus_opus_decipher.txt for decrypting hermetic texts the Magnus Opus style.
The version that does only work with known symbol meanings is prompt_magnus_opus_decipher_near_to_perfect.txt
For texts with lost meanings of the symbols, use prompt_lost_symbol_meanings.txt
---
Deciphering hermetic texts is basically done by translating archetypes in their context according to their context into readable language, followed by updating their original meaning that results from this in a loop until convergence has been reached.
Google Translate helps.
This is EXPERIMENTAL!
License: the Attribution-NonCommercial-ShareAlike 4.0 International
(c) 2026 Dennis Michael Heine