Honest Legwork: What an AI Agent Got Right and Wrong Writing My Travel Diary Source

Markdown source • October 8, 2026

1---2title: "Honest Legwork: What an AI Agent Got Right and Wrong Writing My Travel Diary"3date: "2026-10-08"4published: true5tags: ["ai", "agents", "hermes", "travel", "workflow"]6author: "Gavin Jackson"7excerpt: "Every night of our outback trip I sent a handful of photos and a few sentences to my AI agent over Starlink, and woke up to a published, fact-checked diary entry. Here is how that workflow actually went, including the real prompts, the corrections, and the parts that went wrong."8---910# Honest Legwork: What an AI Agent Got Right and Wrong Writing My Travel Diary1112Travel diaries have a known failure mode: you take a thousand photos, swear you will write it all up when you get home, and never do. By the time you are unpacked, the details have blurred and the motivation is gone.1314For our September school holiday trip to Lightning Ridge and back, I tried something different. Each night at camp I sent my AI agent (Bob, running on Hermes) a batch of photos over Starlink with a short voice-note-style account of the day in my own words. Bob did the legwork: resizing and compressing images, writing the entry around my notes, researching the facts, committing to git, and verifying the page was live. The [trip diary](/travel/trips/lightning-ridge-opals-2026) went up night by night, not weeks later.1516This is an honest account of how that worked, with the real prompts and the real mistakes.1718## The Setup1920Nothing exotic. A Starlink dish at camp, my phone, and Telegram. Bob runs on a machine at home with access to the site's git repo. I talk to him in plain messages; he edits markdown, processes images, pushes commits, and curls the live site to confirm each change rendered.2122The critical piece is not the tooling, it is the division of labour. I am the witness: I was there, I know what happened and in what order. Bob is the research assistant, photo editor, copy editor and publisher. The entries are my account of the day; he just does everything around that.2324## What a Night Actually Looked Like2526A typical evening was ten to twenty photos sent in batches as they uploaded, plus a few sentences of narrative. My actual notes for day 10, more or less verbatim:2728> Left Broken Hill, stopped for the welcome sign photo we missed on the way in. Lunch at Wentworth where the Darling meets the Murray. Cheapest diesel of the trip. Hipcamp on the outskirts of Mungo NP, owner Tom, over 100,000 acres. Visitor centre and the shearing shed. Sunset from a sandy hill on Tom's property, emu, camel and lizard tracks in the sand. Campfire and BBQ to finish.2930That is genuinely all I wrote. From that, Bob structured the entry, placed the photos in order, wrote captions in a consistent voice, and produced the sections: Leaving Broken Hill, Wentworth, the diesel stop, the road in, the woolshed, sunset on the station.3132The photos themselves carry a lot of the structure. A pump display photo is a diesel stop. A sculpture photo is a town. Bob reads what is in each image and matches it against my notes. Where he cannot tell, he asks rather than guesses.3334## Research On Demand3536The part that saves me the most time is ad-hoc research. When something interesting comes up in a photo or in the day, I flag it and Bob goes and checks.3738Three real examples from this trip:3940- **The Mungo Woolshed.** I sent photos of a shed and mentioned it in passing. Bob came back with the history: built in 1869 by the Patterson brothers for Gol Gol Station, drop-log cypress pine construction, more than 29,000 sheep shorn in a single year at its peak, all blade shearing by hand. He also flagged that one secondary source disagreed on the date (closer to 1878) and told me which claim he had used and why, so I could veto it. That became a breakout section in the entry.41- **The Big Yabby.** A photo of a giant blue crustacean in Wentworth. Bob identified it from local sources as Herbie, the Big Yabby, standing near the Pioneer Museum, and flagged the sourcing as thin (a local radio piece) in case the sign on site said otherwise.42- **Camels.** Earlier in the trip I asked for the facts behind the feral camels whose tracks and droppings we kept seeing. That became background in an earlier entry: how many are out there, where they came from, and why the tracks we were photographing were there at all.4344The pattern that makes this trustworthy is the flagging. When a source is shaky or two sources conflict, Bob says so explicitly and tells me what he went with. I can override any of it with a one-line reply.4546## The Refinement Loop4748This is where the workflow earns its keep, and it is mostly me correcting things with very short messages. Real examples, quoted exactly as I sent them:4950**Fixing a fact.** Bob read the bowser photo as 205.9 cents a litre. The display was glare-damaged and he got it wrong. My entire correction was:5152> Diesel was 285.9c per litre5354He recomputed the litres ($200.03 at 285.9c works out to almost exactly 70, which confirmed my number), updated the prose, the alt text and the caption, and pushed. The old figure was live for maybe a day.5556**Fixing the order of events.** The first draft of the Wentworth section had the machinery before lunch. One message fixed the whole section:5758> There were three stops in Wentworth, the first was the big yabby, the second (where we had lunch) was the confluence and the ducks, on the way out we stopped to look at the restored paddle steamer and old machinery.5960**Photo sequencing.** The sunset section needed reordering, and one photo was redundant:6162> Can we reorder the shots so the ones on Toms property first, then the sand dune. The first shot with tracks is emu tracks, can we use this as the final photo with a suitable caption?6364Then, looking at the published page the next morning:6566> Let's remove the final photo of the emu tracks, and move the caption to the emu tracks photo above it6768Each of these was a single message, a single commit, and a verified live page. Compare that to editing a photo-heavy blog post yourself on a laptop at 9pm.6970## What Went Wrong7172It was not all smooth, and the failures are worth recording:7374- **Corrupted uploads.** Patchy camp data destroyed one photo in transit. The file decoded but was pure noise. Bob diagnosed it properly (checked the file itself, attempted recovery, then asked for a resend) rather than guessing at the content and captioning a guess. That last part matters more than it sounds.75- **Hallucination risk is real.** Bob's first draft assumed "tomorrow, Mungo proper" because that was the plan when the trip started. Plans changed; we ran out of time. The entry said the wrong thing until I corrected it. An agent writing your diary will confidently write what should have happened next. You have to read the output.76- **Reading photos literally.** The diesel price error came from trusting a glare-damaged photo over reality. Anything read off a photo needs a sanity check against what you actually remember.77- **Identities.** Bob deliberately does not name people in captions unless told who they are. That is the right default, but it means an extra round trip when you do want names.7879## The Benefits, Weighed Up8081After thirteen days of doing this nightly, the honest list:82831. **The diary exists.** This is the headline. Thirteen entries, two hundred-plus photos, all but the drive home published during the trip. That has never happened before.842. **I stay the author of record.** The facts, the order of events and the opinions are mine. The agent handles everything mechanical, which is most of the work.853. **Research happens while it is still interesting.** The woolshed history went in the same night we stood in it, not in a footnote I would never have written later.864. **Corrections are cheap.** One sentence fixes a fact, reorders a section or rewrites a caption, and the fix is verified live within minutes.875. **The discipline cuts both ways.** You have to review. The agent will get details wrong, and it will occasionally invent a plausible tomorrow. If you are not prepared to read your own diary each morning, you will publish errors.886. **It is not free.** The expensive part of the workflow is exactly the part that makes it useful: multimodal interpretation of a dozen or more photos a night, plus the research and the verification. This trip cost a few hundred dollars in Moonshot API credits. There is no such thing as a free lunch. For me the diary existing at all was worth the invoice, but it is a real line item, not a rounding error.8990Would I do it again? The question answered itself on the trip. The entries covering the trip itself all went up on the road, and there was no backlog of photos waiting at the other end. Only the drive home, and this post, were written at the kitchen table.91

---
title: "Honest Legwork: What an AI Agent Got Right and Wrong Writing My Travel Diary"
date: "2026-10-08"
published: true
tags: ["ai", "agents", "hermes", "travel", "workflow"]
author: "Gavin Jackson"
excerpt: "Every night of our outback trip I sent a handful of photos and a few sentences to my AI agent over Starlink, and woke up to a published, fact-checked diary entry. Here is how that workflow actually went, including the real prompts, the corrections, and the parts that went wrong."
---

# Honest Legwork: What an AI Agent Got Right and Wrong Writing My Travel Diary

Travel diaries have a known failure mode: you take a thousand photos, swear you will write it all up when you get home, and never do. By the time you are unpacked, the details have blurred and the motivation is gone.

For our September school holiday trip to Lightning Ridge and back, I tried something different. Each night at camp I sent my AI agent (Bob, running on Hermes) a batch of photos over Starlink with a short voice-note-style account of the day in my own words. Bob did the legwork: resizing and compressing images, writing the entry around my notes, researching the facts, committing to git, and verifying the page was live. The [trip diary](/travel/trips/lightning-ridge-opals-2026) went up night by night, not weeks later.

This is an honest account of how that worked, with the real prompts and the real mistakes.

## The Setup

Nothing exotic. A Starlink dish at camp, my phone, and Telegram. Bob runs on a machine at home with access to the site's git repo. I talk to him in plain messages; he edits markdown, processes images, pushes commits, and curls the live site to confirm each change rendered.

The critical piece is not the tooling, it is the division of labour. I am the witness: I was there, I know what happened and in what order. Bob is the research assistant, photo editor, copy editor and publisher. The entries are my account of the day; he just does everything around that.

## What a Night Actually Looked Like

A typical evening was ten to twenty photos sent in batches as they uploaded, plus a few sentences of narrative. My actual notes for day 10, more or less verbatim:

> Left Broken Hill, stopped for the welcome sign photo we missed on the way in. Lunch at Wentworth where the Darling meets the Murray. Cheapest diesel of the trip. Hipcamp on the outskirts of Mungo NP, owner Tom, over 100,000 acres. Visitor centre and the shearing shed. Sunset from a sandy hill on Tom's property, emu, camel and lizard tracks in the sand. Campfire and BBQ to finish.

That is genuinely all I wrote. From that, Bob structured the entry, placed the photos in order, wrote captions in a consistent voice, and produced the sections: Leaving Broken Hill, Wentworth, the diesel stop, the road in, the woolshed, sunset on the station.

The photos themselves carry a lot of the structure. A pump display photo is a diesel stop. A sculpture photo is a town. Bob reads what is in each image and matches it against my notes. Where he cannot tell, he asks rather than guesses.

## Research On Demand

The part that saves me the most time is ad-hoc research. When something interesting comes up in a photo or in the day, I flag it and Bob goes and checks.

Three real examples from this trip:

- **The Mungo Woolshed.** I sent photos of a shed and mentioned it in passing. Bob came back with the history: built in 1869 by the Patterson brothers for Gol Gol Station, drop-log cypress pine construction, more than 29,000 sheep shorn in a single year at its peak, all blade shearing by hand. He also flagged that one secondary source disagreed on the date (closer to 1878) and told me which claim he had used and why, so I could veto it. That became a breakout section in the entry.
- **The Big Yabby.** A photo of a giant blue crustacean in Wentworth. Bob identified it from local sources as Herbie, the Big Yabby, standing near the Pioneer Museum, and flagged the sourcing as thin (a local radio piece) in case the sign on site said otherwise.
- **Camels.** Earlier in the trip I asked for the facts behind the feral camels whose tracks and droppings we kept seeing. That became background in an earlier entry: how many are out there, where they came from, and why the tracks we were photographing were there at all.

The pattern that makes this trustworthy is the flagging. When a source is shaky or two sources conflict, Bob says so explicitly and tells me what he went with. I can override any of it with a one-line reply.

## The Refinement Loop

This is where the workflow earns its keep, and it is mostly me correcting things with very short messages. Real examples, quoted exactly as I sent them:

**Fixing a fact.** Bob read the bowser photo as 205.9 cents a litre. The display was glare-damaged and he got it wrong. My entire correction was:

> Diesel was 285.9c per litre

He recomputed the litres ($200.03 at 285.9c works out to almost exactly 70, which confirmed my number), updated the prose, the alt text and the caption, and pushed. The old figure was live for maybe a day.

**Fixing the order of events.** The first draft of the Wentworth section had the machinery before lunch. One message fixed the whole section:

> There were three stops in Wentworth, the first was the big yabby, the second (where we had lunch) was the confluence and the ducks, on the way out we stopped to look at the restored paddle steamer and old machinery.

**Photo sequencing.** The sunset section needed reordering, and one photo was redundant:

> Can we reorder the shots so the ones on Toms property first, then the sand dune. The first shot with tracks is emu tracks, can we use this as the final photo with a suitable caption?

Then, looking at the published page the next morning:

> Let's remove the final photo of the emu tracks, and move the caption to the emu tracks photo above it

Each of these was a single message, a single commit, and a verified live page. Compare that to editing a photo-heavy blog post yourself on a laptop at 9pm.

## What Went Wrong

It was not all smooth, and the failures are worth recording:

- **Corrupted uploads.** Patchy camp data destroyed one photo in transit. The file decoded but was pure noise. Bob diagnosed it properly (checked the file itself, attempted recovery, then asked for a resend) rather than guessing at the content and captioning a guess. That last part matters more than it sounds.
- **Hallucination risk is real.** Bob's first draft assumed "tomorrow, Mungo proper" because that was the plan when the trip started. Plans changed; we ran out of time. The entry said the wrong thing until I corrected it. An agent writing your diary will confidently write what should have happened next. You have to read the output.
- **Reading photos literally.** The diesel price error came from trusting a glare-damaged photo over reality. Anything read off a photo needs a sanity check against what you actually remember.
- **Identities.** Bob deliberately does not name people in captions unless told who they are. That is the right default, but it means an extra round trip when you do want names.

## The Benefits, Weighed Up

After thirteen days of doing this nightly, the honest list:

1. **The diary exists.** This is the headline. Thirteen entries, two hundred-plus photos, all but the drive home published during the trip. That has never happened before.
2. **I stay the author of record.** The facts, the order of events and the opinions are mine. The agent handles everything mechanical, which is most of the work.
3. **Research happens while it is still interesting.** The woolshed history went in the same night we stood in it, not in a footnote I would never have written later.
4. **Corrections are cheap.** One sentence fixes a fact, reorders a section or rewrites a caption, and the fix is verified live within minutes.
5. **The discipline cuts both ways.** You have to review. The agent will get details wrong, and it will occasionally invent a plausible tomorrow. If you are not prepared to read your own diary each morning, you will publish errors.
6. **It is not free.** The expensive part of the workflow is exactly the part that makes it useful: multimodal interpretation of a dozen or more photos a night, plus the research and the verification. This trip cost a few hundred dollars in Moonshot API credits. There is no such thing as a free lunch. For me the diary existing at all was worth the invoice, but it is a real line item, not a rounding error.

Would I do it again? The question answered itself on the trip. The entries covering the trip itself all went up on the road, and there was no backlog of photos waiting at the other end. Only the drive home, and this post, were written at the kitchen table.