Promotional explainer for a Claude Code document-extraction pipeline
Promotional explainer for a Claude Code document-extraction pipeline is a 32.8-second Claude Opus motion design reference by @thedelost, built around code. [Note: the creator shared a partial prompt. Copy this product launch motion prompt, keep the creator's frame counts and palette, and adapt the subject for your own launch film.
@thedelostView post
Prompt
[Note: the creator shared a partial prompt. Wording below is theirs.]
GOAL:
THIS IS AN ACTUAL F*CKING GOLDMINE
Nobody is talking about the cheapest model in your Claude Code pipeline.
MOTION:
The idea: stop generating text to read a document. XERJ Reader never writes a token, it points at the characters already sitting in your file.
Because nothing is synthesized, there is nothing for it to hallucinate. You get exact character offsets, or a strict No.
You just split the pipeline like this:
Opus 5.5 plans · XERJ Reader extracts · Haiku 5.5 verifies
here's the split:
→ Opus 5.5 keeps the planning and the code synthesis, where a frontier model earns its price
→ raw document dumps go straight to XERJ Reader instead of into your context window
→ Reader resolves the query in 80ms and hands back byte-level span offsets
→ Haiku 5.5 checks the schema format, Opus applies the final patch
the bill is the part that stops you:
→ $0.01 per 1M input tokens
→ $0.00 output, because zero tokens are synthesized
→ 261x cheaper than Opus 5.5 on a 1M invoice extraction simulation, $16.75 against $4,368
Opus plans. Reader extracts. Haiku verifies.
JEV engineering is the same move one layer down: the forks that need no thinker (which field, which doc, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split
the full tree
> a non-generative neuro-symbolic model built for document intelligence, not a chat model
> it scans the raw text stream and returns exact character offsets, or a strict No
> it physically cannot hallucinate, because it never synthesizes a single token
> 26x cheaper than Haiku 5.5, and 300x cheaper on input than Sonnet 5.5
> 80ms batch latency instead of multi-second prefill on a 200k token dump
> JEV routes micro-forks in milliseconds
paste the tree and this prompt into Claude Code ↓
Rebuild my extraction path around this tree:
1. Keep Opus 5.5 on planning and code synthesis only, and stop piping raw documents into its context.
2. Route every extraction call to XERJ Reader and take back character offsets instead of prose.
3. Have Haiku 5.5 verify the schema against those spans before Opus touches anything.
4. Log the input cost of every extraction so I can see what the old path was charging me.
5. Route fast deterministic forks (which field, which doc, retries) through JEV.
Show me every change as a diff first. No edits until I say go.
Author original
THIS IS AN ACTUAL F*CKING GOLDMINE
Nobody is talking about the cheapest model in your Claude Code pipeline.
The idea: stop generating text to read a document. XERJ Reader never writes a token, it points at the characters already sitting in your file.
Because nothing is synthesized, there is nothing for it to hallucinate. You get exact character offsets, or a strict No.
You just split the pipeline like this:
Opus 5.5 plans · XERJ Reader extracts · Haiku 5.5 verifies
here's the split:
→ Opus 5.5 keeps the planning and the code synthesis, where a frontier model earns its price
→ raw document dumps go straight to XERJ Reader instead of into your context window
→ Reader resolves the query in 80ms and hands back byte-level span offsets
→ Haiku 5.5 checks the schema format, Opus applies the final patch
the bill is the part that stops you:
→ $0.01 per 1M input tokens
→ $0.00 output, because zero tokens are synthesized
→ 261x cheaper than Opus 5.5 on a 1M invoice extraction simulation, $16.75 against $4,368
Opus plans. Reader extracts. Haiku verifies.
JEV engineering is the same move one layer down: the forks that need no thinker (which field, which doc, retry or stop) go to JEV in milliseconds, and the big models only see the ones that split
the full tree
> a non-generative neuro-symbolic model built for document intelligence, not a chat model
> it scans the raw text stream and returns exact character offsets, or a strict No
> it physically cannot hallucinate, because it never synthesizes a single token
> 26x cheaper than Haiku 5.5, and 300x cheaper on input than Sonnet 5.5
> 80ms batch latency instead of multi-second prefill on a 200k token dump
> JEV routes micro-forks in milliseconds
paste the tree and this prompt into Claude Code ↓
Rebuild my extraction path around this tree:
1. Keep Opus 5.5 on planning and code synthesis only, and stop piping raw documents into its context.
2. Route every extraction call to XERJ Reader and take back character offsets instead of prose.
3. Have Haiku 5.5 verify the schema against those spans before Opus touches anything.
4. Log the input cost of every extraction so I can see what the old path was charging me.
5. Route fast deterministic forks (which field, which doc, retries) through JEV.
Show me every change as a diff first. No edits until I say go.
Questions
What is this product launch prompt for?
It is a Claude Opus motion design brief for “Promotional explainer for a Claude Code document-extraction pipeline”. Paste it into Claude and keep the timings already written in the prompt.
Where is the original film?
The creator is @thedelost. Use View post to open the X video this prompt belongs to.