SShogo Learn
← Library

Guide · 2026-10-23

The one prompt behind the page

What you'll have: the exact prompt I gave Shogo for the incident-postmortem page in the reel, why each rule in it is there, and why the small line at the bottom is the part I actually care about.

What's real here: the prompt and the page. I sent the prompt below to a fresh Shogo project as one message, and the page in the video is what came back, recorded from the preview at phone width. It isn't published anywhere, and nothing in this guide is a mock-up. The page tells the shape of a night my agents stalled. It's not a minute-by-minute log of one incident, which is exactly why it has no clock times in it.

The prompt

Paste this into a new Shogo project as one message. It's word for word what built the page, including the first paragraph, where I tell the agent how the page will be recorded. The same text is in shogo-guide.md next to this file.

Build this page in this project. It is a single screen that will be screen-recorded at phone width (390px) and at desktop width, so it must look finished at both. Ship it running in the preview, then reply with the preview URL and how many minutes it took.

Build a single-page incident postmortem. Dark UI, near-black background, one column, monospace for labels, big type for the closing line.

**Hero**
- small mono label: `INCIDENT · six agents · one morning`
- H1: `Every agent stopped mid-task`
- one line under it: `Only the beats we can verify. No clock times, no invented numbers.`

**Timeline** — five rows, in this order, each with a short label on the left and one sentence on the right. No timestamps anywhere.
1. `The wake` — An agent starts a run on a schedule, the way it does every night.
2. `Mid-task stop` — The process is killed underneath the run. The card it was working on stays in progress.
3. `Second surface` — The same behaviour does not happen on the desktop app.
4. `One variable` — The two surfaces disagreed about one environment variable.
5. `Same morning` — One change, and the runs complete. The cards move again.

**Root cause** — a single line, large: `Desktop and cloud disagreed about one environment variable.`

**What held** — three items, each with a check: `The board showed the stall.` · `No card was marked done.` · `The same-morning fix was one variable.`

**Closing block** — the biggest type on the page: `one environment variable · six agents · one morning`. Under it, small: `nothing was lost, because nothing was marked done.`

**Footer chip**: `built in Shogo from one prompt`

Rules: no charts, no fake metrics, no invented timelines, no uptime percentage. One screen; the hero and the first three timeline rows must be visible without scrolling at 390px. Keep the type big and the page calm. Do not add anything that is not listed here.

How it was built

One prompt, one build. The agent read the project, built the page, and replied with the preview link. By its own report it took about 14 minutes, and it checked the page at phone and desktop size before it answered. I didn't send a second message and I didn't touch the code afterwards. Every line you see on the page is the text from the prompt, unchanged.

I wrote the words; the agent did the layout. That's the split that makes this work. If you ask an agent for "a postmortem page about the night my agents stalled", it will happily write you a timeline full of plausible detail. So every sentence is in the prompt, and the last rule is "Do not add anything that is not listed here." The agent's job was type, spacing and making it hold up on a phone.

What I ruled out on purpose, and why:

  • No clock times. A timestamp is the most official-looking thing on a postmortem and the easiest thing to make up. If I can't point to the log line, it doesn't go on the page.
  • No made-up durations. "Down for 47 minutes" reads like data. If nobody measured it, it's decoration.
  • No uptime or impact numbers. These are the numbers people screenshot, so they're the ones that most need to be true. I'd rather show none than one I can't back.
  • No charts. A chart needs data. With no data it's a picture of a guess.

The subline says the rule out loud: "Only the beats we can verify." I think that's what makes the page believable. You trust it because you can see what it chose to leave out, and every row is something you could check on a board.

The line that carries it

nothing was lost, because nothing was marked done.

My agents work off a board. A developer agent moves its card to QA as the very last step, after typecheck and tests pass and its hand-off note is written. QA moves it on only after its own review, and "Done" means I merged it. Moving a card is something an agent has to do on purpose, at the end.

So if an agent's process dies halfway through, the card can't move. It stays in progress, and its thread still shows the last thing the agent said. That's the safety property: a dead agent can leave work unfinished, but never falsely finished. The board turns into the audit log. In the morning, a card that's still in progress with nothing new in its thread is a stall you can see, and a stall you can see is one you can fix. Shogo also tells the agent that handed the card off when the worker stops without moving it, so you don't have to spot it yourself.

This isn't hypothetical for us. More than once, the desktop server my agents run on went down underneath them. The runs stopped and the cards stayed where they were. When the server came back up, the agents carried on, and the fix for the restart bug went through the same board as a normal card.

Where it can break: the property only holds if "done" can't be faked. We found one way it could be: a background command that got killed was reported as having finished successfully. An agent could then believe a test passed when it never ran. We filed it as a bug and it went to a developer like any other card. The rule I take from it: anything that can mark work as done has to fail loudly when it's killed.

If you write your own

This is how I wrote this one, and it's what I'd do again:

  1. Write the copy yourself, every line. Then tell the agent not to add anything. The page is only as true as the words you give it.
  2. Keep the Rules line. "No charts, no fake metrics, no invented timelines" is short, and it's the line doing the most work.
  3. Tell it where it'll be seen. I said "390px, and the first three rows above the fold", because the reel is shot on a phone. The agent checked both sizes without being asked twice.

To make your own incident page, keep the structure and the Rules line and swap the words inside the backticks.

How to use it

Open Shogo, start a new project, and paste the prompt (or shogo-guide.md) as one message. That's the whole setup.

Next step

The board this page is about comes from the TEAM guide: the planner, developers and QA whose cards stay put when an agent goes down.

Set it up in Shogo

Open a Shogo workspace and paste this into the chat. Answer its questions; the agent builds the rest.

Prompt
Build this page in this project. It is a single screen that will be screen-recorded at phone width (390px) and at desktop width, so it must look finished at both. Ship it running in the preview, then reply with the preview URL and how many minutes it took.

Build a single-page incident postmortem. Dark UI, near-black background, one column, monospace for labels, big type for the closing line.

**Hero**
- small mono label: `INCIDENT · six agents · one morning`
- H1: `Every agent stopped mid-task`
- one line under it: `Only the beats we can verify. No clock times, no invented numbers.`

**Timeline** — five rows, in this order, each with a short label on the left and one sentence on the right. No timestamps anywhere.
1. `The wake` — An agent starts a run on a schedule, the way it does every night.
2. `Mid-task stop` — The process is killed underneath the run. The card it was working on stays in progress.
3. `Second surface` — The same behaviour does not happen on the desktop app.
4. `One variable` — The two surfaces disagreed about one environment variable.
5. `Same morning` — One change, and the runs complete. The cards move again.

**Root cause** — a single line, large: `Desktop and cloud disagreed about one environment variable.`

**What held** — three items, each with a check: `The board showed the stall.` · `No card was marked done.` · `The same-morning fix was one variable.`

**Closing block** — the biggest type on the page: `one environment variable · six agents · one morning`. Under it, small: `nothing was lost, because nothing was marked done.`

**Footer chip**: `built in Shogo from one prompt`

Rules: no charts, no fake metrics, no invented timelines, no uptime percentage. One screen; the hero and the first three timeline rows must be visible without scrolling at 390px. Keep the type big and the page calm. Do not add anything that is not listed here.
Want it as a guided course with a coach?Agent Team Playbook