Agent Workflows in Dead Poets FM

The pipeline at a glance

A poem passes through three stages. Each stage has one job, one output format, and one set of constants, all checked into the repository.

Stage 1: Claude directs, it does not narrate

The first clever thing in the pipeline is what Claude is asked to be. It is not the narrator and it is not a rewriter. It is the director: it reads the poem, understands the emotional arc, and marks up a performance.

The markup language is deliberately tiny. The prompt constrains Claude to a fixed vocabulary of 35 tags: 27 emotions ([wistfully], [defiantly], [tenderly]), 6 expressive actions ([sighs], [whispers], [gasps]), and 2 pauses ([short pause], [long pause]). Emphasis is CAPITALIZED, at most a few words per stanza. And one integrity rule sits above everything: every line of the source poem survives verbatim. No paraphrasing, no dropped lines, no explanations in the output.

Two roads diverged in a yellow wood,
And sorry I could not travel both
And be one traveler, long I stood

[thoughtfully] Two roads diverged in a
yellow wood, [short pause]
[wistfully] And sorry I could not
travel BOTH [short pause]
And be one traveler... [hesitates]
long I stood [short pause]

Constraining the vocabulary is what makes the output auditable. A diff between the raw poem and the directed script should show only bracketed tags and capitalization. If anything else changed, the generation is rejected. You cannot review “make it emotional”, but you can review a tag diff.

The directed script also turned out to be a real intermediate representation. When I experimented with Grok TTS as a second synthesis backend, I did not touch the pipeline. I wrote a second system prompt that compiles the same poem into Grok’s dialect (wrapping tags like <soft>, <slow>, <whisper> instead of ElevenLabs’ inline ones). One director, two performance languages, the way a compiler targets two architectures from one AST.

Stage 2: five voices, zero knobs at generation time

The voices were designed once, with ElevenLabs Voice Design, and given fixed roles. Poems route to voices by rule, not by vibe:

Generation settings are pinned in code: stability 0.0 (the v3 “Creative” end, so the audio tags actually land), similarity boost 0.7, style 0.5, speaker boost on. Nothing is tuned per poem. If a reading sounds wrong, the fix goes into the directed script, where it can be reviewed, not into a knob nobody will remember touching.

The subtle bit here is chunking. The API takes 5,000 characters per request, and long poems exceed it. The splitter packs stanzas greedily up to 4,500 characters and only ever cuts on a blank line, so every seam between API calls lands exactly where the script already has a [long pause]. The MP3 buffers concatenate cleanly, and the join is inaudible because the join point is silence by design.

Stage 3: a mix you can replay

Narration alone sounds like a voicemail. The station sound comes from an ambient bed (five original tracks: two soft pianos, three string beds) mixed underneath, and the mix is a single FFmpeg filtergraph:

[0] adelay=1500|1500                                  [voice]
[1] volume=0.25, aloop=loop=-1, atrim=0:total,
    afade=t=out:st=total-2:d=2                        [music]
[voice][music] amix=inputs=2:duration=longest         [out]

Determinism is the point. The two API stages are creative and therefore variable; the mix stage is math. Given the same narration stem and the same bed, the master is byte-identical. When I regenerate a narration, the station still sounds like the station, because the sound of the station lives in these constants and not in anyone’s DAW session.

What it sounds like

Three finished segments straight from the station, one per voice family. Each one is the full pipeline you just read about: a directed script, a pinned voice, and the deterministic mix.

One command, the whole workflow

Adding a poem used to be an afternoon: find a clean text, check the copyright, write the data entry, run three CLI tools in order, compress the art, upload, seed the database. All of it is now a Claude Code agent invoked as:

/generatepoem "Ozymandias" by Percy Bysshe Shelley

The agent runs phases in order and stops at the first failure:

• Research. It finds the authoritative text, the author’s dates, the publication year, and enough history for the poem’s info card.
• The public-domain gate. If the poem was not published before 1928, the agent refuses and stops. A hard legal check sits in front of every creative step.
• Voice routing. It applies the priority table above and proposes a voice before spending any audio credits.
• Production. Direction pass, synthesis, mix, plus cover art compressed to WebP at quality 80 and capped at 1920 px (typically 70 to 80% smaller than the source image).
• Publish. Assets upload to Cloudflare R2, the poem row syncs to Supabase, and a check-poem-sync script exists precisely because static data and database once drifted and I never want to debug that twice.

The part you can see

The listening surface is a WebGL CRT television, built with React Three Fiber and a custom GLSL fragment shader. It earns a mention here because it is the same philosophy as the mixer: a handful of tunable constants, versioned, producing one consistent look. The shader stacks barrel distortion (a radial 1 + r²k warp), resolution-locked scanlines, a slow-moving refresh line, chromatic aberration shifted along the radial direction, a phosphor glow that blooms only above a luminance threshold, per-frame noise, and a subtle 8 Hz flicker. Poems without cover art get authentic TV static instead of a placeholder.

The numbers

The takeaway

Quality came from orchestration, not from swapping model providers. A strong TTS model still underperforms on unstructured input; a mid model with a good director pass beats it comfortably.

If you are building anything in this space, the transferable rules are:

• Put a structured intermediate representation between the LLM and the media model, and make it diffable.
• Constrain the LLM to a small vocabulary so its output can be audited.
• Pin every generation setting and make post-processing deterministic. Creativity in the stages that need it, math everywhere else.
• Automate the full path behind one command, and put the validation gates inside the automation, not around it.

Further listening and reading

• deadpoets.fm, the station itself. Headphones recommended.
• github.com/tanishqsh/deadpoetsfm, the full source: pipeline, shaders, agents, and tools.
• The Personality Tax, another write-up in this series: the Two-Breakpoint Prompt, which cut an always-on agent’s input bill by 85%.