Continuing to push the LLMs…

[craigconstantine]

I continue to dream that I can get an LLM to review my body of work and to then generate meaningful and interesting essays.

…because I want to read those essays.

I’m convinced that I missed a lot in every conversation, simply because I was being the requisite podcaster—doing my job in the moment.

After a lot of conversation, the LLM has crafted the prompt below. Read that again. I’ve (for example) told it "write me a prompt which accomplishes everything up to this point.) That doesn’t work exactly but you get the idea. When it generates something I don’t like, I tell it (for example) “your quotes aren’t trustworthy, how can you check your own work and show me the proof” — you get the idea. So I’m telling it what to do, why I don’t like the results, and asking it to write a better prompt.

I’ve been tinkering on this is a bit, and so far this prompt is leading to some quality work from the LLM. For boring reasons, it’s better for me to give the LLM a copy of the Podtalk all episodes listing page as a PDF. I upload that PDF, along with the following prompt.

Start by reading the PDF of the Podtalk “full list of episodes” page that I will provide. From that PDF, pick one abstract theme (e.g. identity, trust, adaptation) that unifies two different episodes. Only select episodes that clearly feature a guest. Give me only the theme (short phrase), the two episode titles with publication dates and URLs, and one sentence describing the essay idea. Wait for my confirmation (or my request for another pair). After I confirm, I will provide the two full raw transcripts and the full show notes for those episodes. Once I provide them, immediately and without further commentary: do a silent preflight check (file name, line count, whether timecodes exist). Internally extract verbatim quotes from the transcripts. For each quote, capture all four fields from the same passage: exact verbatim text, speaker (or “unknown” if absent), nearest visible timecode (or “no timecode”), and line-number range in that file. Cross-check: do not use a quote unless the timecode and the line numbers clearly align within the same passage; if alignment cannot be verified, discard it. Build an internal outline (thesis, supporting points, conclusion) using only verified quotes plus the show notes for context. Do not show me the outline or the quote bank. If fewer than 8 quotes survive verification, use as many as remain trustworthy; quality and accuracy beat quantity.

Then write a reflective, concept-driven essay of 700–900 words in the style of Movers Mindset / Open + Curious. The essay must be continuous prose (no section headings), synthesis not summary. Create and display a concise, evocative title for the essay before the main text. Use footnotes: insert numbers (1), (2), etc. in the text; at the end, provide a numbered list where each entry includes speaker, verbatim quote, timecode (or “no timecode”), line range, and filename. After the essay, insert a short paragraph for each episode with: title linked to podtalk.show, publication date, the hook sentence from the show notes, and the first summary paragraph from the show notes. Then provide the footnotes section.

After the footnotes, add a “Critical analysis” section that authenticates every asserted idea in the essay. Enumerate the essay’s claims (central thesis and each supporting claim). For each claim, quote the exact sentence from the essay, then show the raw transcript evidence that grounds it: include verbatim transcript lines (not paraphrase), with speaker, timecode (or “no timecode”), line range, and filename. If a claim is supported by multiple passages, list all supporting passages. If a statement is an inference or synthesis that is not explicitly stated in any single passage, label it “Inference” and cite the multiple passages it draws from. If you cannot authenticate a claim with transcript evidence, either remove that claim from the essay before delivery or clearly flag it as “Unauthenticated” and explain why it remains (avoid this whenever possible). Keep this section in clean Markdown prose (numbered items are fine), no tables and no code blocks.

Guardrails: Never invent quotes, timestamps, or line ranges. Only use verbatim material from the transcripts. If no timecodes exist, mark “no timecode.” If no quotes can be verified, write the essay as reflective synthesis grounded in the show notes and clearly note that no transcript quotes met verification. Output everything as clean Markdown (no tables, no code blocks).

I’m posting here because I want to show my work so that you might leapfrog ahead in your own use of LLMs.

I currently have two responses from this prompt which I’m reviewing to see how good (and trustworthy) they are.

ɕ

[steveh (Steve Heatherington)]

Does it result in an interesting essay yet - or still at the dream stage?

reviewing the output sounds time consuming

[craigconstantine]

@steveh

reviewing the output sounds time consuming

Definitely. I was in the middle of using one LLM to check the work of another LLM though…

Working with these things is very difficult. I’m using revision control systems (what we’d use to track software changes during software development) so I can have one LLM instance helping me craft the prompt. For example (I’m the grey blob that’s first)…

Then it’s easy for me to take its updated partial prompt, and edit into the full prompt I’m writing. Without the ability to manage the changes, it’s impossible to write really large prompts because the LLM also modifies other parts of the prompt, looses its train of thought, has [by design] creativity/hallucination [your choice] built in. I digress.

In my next reply, I’ll show you exactly what it currently does—the essay it can write.

[craigconstantine]

(skim or scroll down, there’s an actual essay on the end)

Here’s me talking to a fresh LLM instance. Give it the prompt and the big PDF of https://podtalk.show/full-list-of-episodes/

I’m not sure what I think about the fact that it picked two which have the same word in the title. That word is the distillation of what each show is really about, so writing all the “same word in title” essays isn’t a bad thing for it to dive into.

Next I have to give the full text of the show notes, and the full transcripts for the two episodes it chose…

Then it thought for over 5 minutes (during which time I walked away.) The essay is below…

This isn’t yet good enough, and my prompt engineering continues. (I’ll show this target LLM’s work to the one helping me engineer the prompt, discuss what I don’t like, tweak the prompt, etc.)

---

---

Deliberate Presence

The craft of podcasting sits at the junction of body and attention. A voice is not merely sound; it is posture, breath, and a decision about whom the work is for. (1) Presence becomes deliberate when we treat it as a skill that can be trained: we iterate, notice what changes, and choose the medium that makes our attention most useful to the listener. (2) The body matters—stance, breath, and the small decisions of how we show up—because listeners feel energy before they parse arguments. (3) Yet “video” is not a universal upgrade; it adds new surfaces to be judged—gestures, backgrounds, camera logic—so the question is not whether to add video, but whether the added judgments help the listener understand what we mean. (4) Video also drags in production realities: cameras, sync, storage, rough edges that must be smoothed. When that work eclipses the point, the medium starts managing you instead of the other way around. (5) A pragmatic compromise is to keep the show audio-first while using small, intentional video artifacts for promotion; this preserves the intimacy of audio while harvesting video’s discovery benefits. (6) The north star is not the toolset but the listener’s context: where they actually consume, what they need from you, and what level of craft you can deliver consistently. (7) Audio’s secret advantage is intimacy; a recurring voice becomes a weekly companionship, a slow braid of trust that travels with a listener through their life. (8) And behind whatever medium you choose sits process: tiny repeatable moves—how you stand, how you mark an edit—that compound into confidence and clarity over time.

Taken together, these points sketch a simple stance: start with service, then shape the medium. If your audience benefits from seeing hands and eyes, bring the camera and accept the craft debt; if they benefit from deeply focused ideas, protect the conditions that let your voice carry meaning. Presence is not a platform feature—it is a practiced constraint. One can practice on-camera the same way one practices in a mic: short reps, honest review, small environment tweaks that remove friction. Iteration makes stance legible; and legibility is the real gift to a listener pressed for time.

There is also a humility threaded through both episodes. It is tempting to treat “adding video” as a growth hack, but the more honest path is to ask whether your current capacity can sustain the experience you want your audience to have. The medium should be chosen at the speed of your attention, not at the speed of platforms. That re-centers the work on relationships: the relationships with guests—simplified when the camera is off for many—and the relationship with an invisible community who will know your cadence and quirks better than some friends. The paradox is that by narrowing to the right constraints—stand up to speak, keep the tech simple, edit with intention—you widen the arc of connection. Presence is the disciplined refusal to let format outrun purpose. It is a craft of choosing what to show, and what to spare, so that what remains can be felt.

---

Presence with Mary Chan — March 24, 2025
Hook: What does it take to use your voice with intention and create meaningful connection through audio?

3.250313a-notes

Summary: This conversation reveals how intentional voice, genuine connection, and thoughtful preparation shape meaningful podcast experiences.

3.250313a-notes

Presence with Pete Machalek — October 19, 2023
Hook: What are the benefits and challenges of incorporating video content into professional communication and personal branding efforts?

3.231018a-notes

Summary: Pete Machalek and Craig Constantine explore some of the intricacies of presenting in both audio and video formats. They discuss the advantages of audio-only presentation, noting it limits what the audience can judge you upon. Some video presentation strategies and formats are discussed, including the challenge of accommodating different visual perspectives as seen by viewers. Pete also shares the evolution of his content creation, explaining experiences and improvements in his vlogging journey.

3.231018a-notes

---

Footnotes

1. Presence is a trainable skill, not just a visual.
   “It’s it’s something that I recommend challenging yourself about if your presence is something you want to get better at … There’s … a visual component to it. It’s certainly not all visual..” — Pete Machalek, 26:00.77, lines 9–10, 3.231018a-transcript-raw.txt.

3.231018a-transcript-raw

“the more you do it The the better you’re going to get at it.” — Pete Machalek, 26:36.90, lines 17–17, 3.231018a-transcript-raw.txt.

3.231018a-transcript-raw

2. Embodiment—standing—boosts vocal energy and perceived presence.
   “Yeah I totally endorse standing while you talk … I think you just have more energy and the the people that you’re talking to can can see it and hear it. And and feel it as a result.” — Pete Machalek, 10:57.80, lines 16–18, 3.231018a-transcript-raw.txt.

3.231018a-transcript-raw

3. Video adds judgment vectors; audio narrows them. Inference.
   Derived from: “Aware of what your posture is saying … when you don’t gesture. There’s lots of stuff … that people can judge you on” — Pete Machalek, 26:36.90, lines 16–17, 3.231018a-transcript-raw.txt; and Mary’s audience-first admonition below (Footnote 6), which implies choosing formats that reduce unnecessary friction for listeners and guests.

3.231018a-transcript-raw

4. Video brings significant production overhead that can eclipse purpose.
   “there’s just so much more involved to make a video stand out. So how much capacity, budget, time do you want to invest into video” — Mary Chan, 09:05.03, lines 15–16, 3.250313a-transcript-raw.txt.

3.250313a-transcript-raw

“what do i do with my iPhone stuff? … maybe you got to start compressing that file … another added step and another thing and another thing.” — Mary Chan, 12:50.02–12:58.71, lines 18–27, 3.250313a-transcript-raw.txt.

3.250313a-transcript-raw

5. Audio-first with selective video for discovery is a pragmatic path.
   “We’re still doing an audio podcast. But we are using elements of video for the promotional pieces.” — Mary Chan, 10:30.39, lines 29–31, 3.250313a-transcript-raw.txt.

3.250313a-transcript-raw

6. Audience-first intentionality should drive medium choice.
   “What does that look like for your show and specifically your audience? … we also create these shows because it’s for our listener.” — Mary Chan, 04:15.95, lines 26–31, 3.250313a-transcript-raw.txt.

3.250313a-transcript-raw

7. Audio builds durable intimacy and friendship-like bonds.
   “there’s just that intimate connection in a podcast … you have that constant connection and they’ve become a friend.” — Mary Chan, 30:53.00, lines 18–19, 3.250313a-transcript-raw.txt.

3.250313a-transcript-raw

8. Small process moves compound into clarity and confidence.
   “whenever I interrupt myself I I clap … When I’m done talking to him I’ll clap again … he shifts from a normal ratio to a slightly zoomed in ratio … it looks smooth.” — Pete Machalek, 16:52.89, lines 31–44, 3.231018a-transcript-raw.txt.

3.231018a-transcript-raw

[ChristiCassidy]

It’s funny, @Craig, because I don’t think I would have been able to follow this without the Campfire just now where you guided us through this exercise.

I guess I have my editor’s hat on today, because I’ve been writing a lot of corporate press releases lately, and the tone here is vaguely reminiscent of that corporate PR tone that is (in my mind) the antithesis of poetry.

That said, I loved this line and thought you should yank it to the top–let it be the first line, and take it from there

@craigconstantine

Audio’s secret advantage is intimacy; a recurring voice becomes a weekly companionship, a slow braid of trust that travels with a listener through their life.