Vercel Labs Next Browser Next Browser

synthesize-transcript

Read the transcript at:

$ARGUMENTS

Then analyze how the agent used next-browser commands. Your goals are to:

1. Find generalizable learnings — things that would help any agent using this tool — and propose targeted additions to SKILL.md.
2. Surface feature requests and bug fixes for the CLI itself — gaps in functionality, incorrect behavior, or misleading output that should be fixed in code.

Process

1. Read the transcript in chunks (it's large). Focus on assistant messages and tool calls/results involving next-browser commands.

2. Identify friction and patterns. Look for:

   • Commands that failed or produced unexpected results
   • Misunderstandings about how a command works
   • Workflows the agent discovered through trial and error
   • Repeated mistakes that guidance would prevent

3. Filter ruthlessly for overfitting. For each candidate learning, ask:

   • Would a different agent on a different project hit this same issue?
   • Is this a property of the tool or a property of this debugging session?
   • Is this already obvious from the command's existing docs?
   • Is this prescribing a specific workflow vs documenting tool behavior?

   Discard anything that is:

   • A workflow pattern specific to one task (e.g., "copy dirs for before/after")
   • Advice an agent could derive from reading existing docs
   • Specific to a particular project, page, or debugging scenario

   Keep as SKILL.md learnings:

   • Genuine tool constraints any agent would hit (e.g., eval quirks, Playwright behavior)
   • Non-obvious failure modes with clear mitigations
   • Command interactions that aren't documented

   Separate out as feature requests or bug fixes (not SKILL.md changes):

   • Workarounds for bugs that should be fixed in code
   • Missing functionality the agent needed and had to hack around
   • Misleading output or docs that don't match actual behavior

4. Present your findings. Organize into two sections:

   SKILL.md learnings — for each, show:

   • What happened in the transcript (brief)
   • The proposed SKILL.md addition (exact text)
   • Why it's generalizable (one sentence)

   Feature requests / bug fixes — for each, show:

   • What happened in the transcript (brief)
   • What the CLI should do instead (proposed behavior)
   • Whether it's a bug fix (current behavior is wrong) or a feature request (new capability needed)

   Ask the user to approve or reject each item before acting.

5. Consider new Scenarios. Beyond command-level learnings, check whether the transcript reveals a scenario not covered in the existing ## Scenarios section of SKILL.md. A scenario earns its place only if it requires domain knowledge an agent wouldn't derive on its own — non-obvious mental models, ordering constraints, or decision frameworks that go beyond "use these commands and compare results."

   Ask yourself: could an agent figure out this workflow just by reading the existing command docs and applying basic debugging instincts? If yes, it's not a scenario — it's just competent tool use. If no — if it requires understanding something about React, Next.js, or the PPR model that isn't in the CLI docs — then it's a candidate.

   Present scenario candidates the same way as command learnings: evidence from the transcript, proposed text, and why an agent couldn't get there alone. Ask for approval before adding.

6. Apply approved changes.

   • SKILL.md learnings → add inline to the relevant command section in the repo-root SKILL.md. Scenarios → ## Scenarios section.
   • Feature requests / bug fixes → we own this package, so implement the fix or feature directly in the codebase. For non-trivial changes, plan the implementation and confirm with the user before writing code. Include the transcript evidence in commit messages or PR descriptions for context.

Anti-patterns

• Don't add "tips" or "best practices" sections. Guidance belongs next to the command it's about.
• Don't add workflow recipes ("first do X, then Y, then Z"). Document tool behavior, not agent strategy.
• Don't inflate existing docs with caveats that rarely apply.
• When in doubt, leave it out. A wrong or noisy addition is worse than a missing one.

Source: vercel-labs/next-browser — distributed by TomeVault.