The agent skill

Hand Claude Code, Codex, Cursor or any agent a folder of recordings. It finds every task, then investigates, reproduces, builds and tests each one.

A hand-off prompt covers one goal. The skill covers the whole job: give your agent a folder from your library and it reads every recording there, separates what you raised into tasks, and works each one through to a tested change, with subagents doing the digging. You record; it builds.

It is an agent skill: two Markdown files that Claude Code, Codex, Cursor, Gemini CLI, OpenCode, GitHub Copilot and any other agent that reads SKILL.md pick up on their own.

Install

If an agent has run on your computer, Startwell Capture offers the skill once, at the top of your Inbox. Click Add the skill and the app keeps it up to date from then on. Settings, Agent access has the same button, and Remove. Or run this once in a terminal on macOS or Linux; run it again to update.

Terminal
curl -fsSL https://capture.startwellventures.com/skills/install.sh | bash

On Windows, in PowerShell:

PowerShell
irm https://capture.startwellventures.com/skills/install.ps1 | iex

Either way the same files land in two folders in your home folder, which between them reach every agent. On Windows they are %USERPROFILE%\.agents\skills and %USERPROFILE%\.claude\skills.

FolderRead by
~/.agents/skills/startwell-captureCodex, Cursor, Gemini CLI, OpenCode, GitHub Copilot and others that share this folder
~/.claude/skills/startwell-captureClaude Code, which reads only its own folder

Prefer to read it first? The skill is served from capture.startwellventures.com/skills/startwell-capture/SKILL.md with briefs.md beside it. To remove it:

Terminal
curl -fsSL https://capture.startwellventures.com/skills/install.sh | bash -s -- --uninstall
PowerShell
& ([scriptblock]::Create((irm https://capture.startwellventures.com/skills/install.ps1))) -Uninstall

Use it

Open your agent in the repo the recordings are about and hand it a folder: a topic, one pack, or the whole library. Any file inside works too, so a pack path from the Copy as menu is enough.

Claude Code, Codex or Cursor
> /startwell-capture "~/Startwell Capture/Demo Shop"

A topic: every pack in it, oldest first. Add "last 3" or "since Monday" at the end to narrow it.

> /startwell-capture "~/Startwell Capture/Demo Shop/2026-11-03 1402 Demo card declined at checkout"

One recording.

> /startwell-capture "~/Startwell Capture/Demo Shop" plan

Stops after the task list and waits for your go-ahead.

You can also just paste the path, or say “work through what I recorded about the checkout”: an agent picks the skill up on its own when it sees a library path.

What happens

  1. Scope. It lists the packs in the folder and checks that the apps and pages in them match the repo it is in. A mismatch stops it before any work.
  2. Triage. It reads every README and issues.md, the transcripts where those leave the meaning unclear, and opens the screenshot at each moment you flagged something. The result is a ledger: one row per task, with your words, the time, and the screenshot as evidence. Repeats are merged, a later recording wins over an earlier one, and ideas and questions go under “Not built” for you.
  3. Each task, in five stages, each a fresh agent with its own brief: investigate the code behind the screenshot; reproduce the problem as a failing test or in the running app with a screenshot; build the smallest change that makes it pass, in one commit; verify it end to end along the path you took, with a screenshot beside your original; review the diff against your words.
  4. Close. It runs the repo’s checks once on the branch, adds a note to the topic when the MCP connection is set up, and reports: what shipped, with commit, test and screenshot; what it could not reproduce; what belongs elsewhere; and what only you can check, such as a permission or the real clipboard.

Bugs go first, then changes. Where the agent can run subagents in parallel it investigates several tasks at once but builds at most two at a time, with one test run at a time, so it stays light on your computer. Work lands on a branch named capture/…; it pushes or opens a pull request only if you ask.

The ledger

The ledger is a Markdown file at .startwell-capture/work/ in the repo, kept out of git. Every stage writes its outcome onto the task’s row as it lands, so a session that stops part way can pick up where it left off, and you can read at any point what is done, what is not, and why.

.startwell-capture/work/2026-11-03-demo-shop.md
| id | kind | task | evidence | status |
|---|---|---|---|---|
| T1 | bug | Postcode with a space is declined at checkout | Demo card declined at checkout 00:42, steps/004.png | shipped a1b2c3d, checkout.test.ts, T1-after.png |
| T2 | change | Show the card error beside the field, not at the top | same, 01:10, steps/007.png | shipped e4f5a6b, …
| T3 | bug | Order history shows the wrong date | 2026-11-03 1420 Orders 00:08 | not reproduced: dates correct on main; needs an order placed before the clock change |

## Not built
- idea: "maybe we could remember the card" (01:30)

What it holds to

  • A task is shipped only after a failing proof turned green and an after screenshot exists.
  • It stays inside what you recorded. Something else it notices becomes a row under Not built.
  • Screenshots and clicks are trusted over the transcript, which is speech-to-text; it goes by your meaning.
  • It never widens the ask, never pushes, and never files anything anywhere unless you say so.

For a chat that cannot open files, use a hand-off prompt with full transcripts instead. The skill needs an agent on the computer that holds the library, or the folder copied there.