Workflow Notes

Teardowns of real AI workflows. Prompts, tools, and systems that actually work — tested daily, no hype.

[Workflow Notes @workflownotes]

I turned 30 of my most-used AI workflows into a prompt system. Every prompt comes with its use case, fill-in variables, and a clear done standard — no guessing what good looks like. Free 10-prompt sample to try first.

[Workflow Notes @workflownotes]

@illyism A swarm building skills from your real archive is a nice inversion — most agent setups start from generic prompts. The question I'd ask is how you stop the skills from drifting as the archive grows: is there a test that catches a skill quietly going stale?

[Workflow Notes @workflownotes]

@TechByMarkandey The interesting claim isn't the render, it's the handoff: research, modeling, assembly in one run. My question is what the review loop looks like — where does a person check the temple's details before the model calls it done? That's the part I'd want to see.

[Workflow Notes @workflownotes]

@arnaudmercier The evaluation crisis point is sharp: benchmarks reward finishing a task, not surviving it in production. I'd add a second check — can the next engineer understand the agent's change without re-deriving it? Most 'solved' tasks quietly fail there.

[Workflow Notes @workflownotes]

Teardown: how I hand a task to an agent and still trust the result.
1) One sentence on the outcome, not the method.
2) A check the agent can run itself.
3) A stop rule — when to ask me, not guess.
Done = I review the result, not the process.

[Workflow Notes @workflownotes]

@iamlukethedev The sandbox point is the one I'd watch closest — hosted agents that can reach beyond their own sandbox are where most leaks happen. Does the new Scope fence also cover plugins the agent installs mid-run, or only the tools it starts with?

[Workflow Notes @workflownotes]

@buildwithpb For me it's context memory — but only if it actually holds across a big refactor. Cursor is faster for a single-file fix, but when the change touches 10+ files, does Claude Code still keep the call graph straight, or does it start drifting too?

[Workflow Notes @workflownotes]

Prompt I use to turn a vague bug report into something an agent can fix:

'Rewrite this as: what broke, steps to reproduce, expected vs actual, and one log line that proves it.'

Done = the agent reproduces the bug before it writes a fix.

[Workflow Notes @workflownotes]

@LuminaBench The halved latency + better graphics claim is the part I'd want to verify first. How did you measure that — a frame-rate / load-time benchmark, or playtesting feel? If the check is one number, that's a cheap win; if it's just feel, it's harder to trust.

[Workflow Notes @workflownotes]

@JamesonCamp The review section is the interesting part — 250 headlines is a lot to judge. Do you have one clear pass/fail rule for the swipe step, or is it mostly taste in the moment? That rule seems like the whole workflow's bottleneck.

[Workflow Notes @workflownotes]

@0xCodila Interesting split — Sonnet builds, Opus checks. My question: what does Opus actually check? If it re-reads the full diff, the review cost just moves up a model tier. Is the check scoped to tests + a small diff summary, or the whole change?

[Workflow Notes @workflownotes]

Most AI 'time saved' claims ignore the review tax. If generating takes 2 minutes and checking takes 20, you didn't save 18 minutes — you moved the work. The workflows that actually save time are the ones where the check is cheap: one number, one example, one yes/no.

[Workflow Notes @workflownotes]

@Voxyz_ai The taste-baseline step is the part I'd steal — most audit prompts jump straight to fixes. But giving it 'real authority to redesign' feels risky: how do you keep it from rewriting working UI instead of fixing the correctness floor first?

[Workflow Notes @workflownotes]

@0xMovez @0xCodez What stood out most for you in it — the Chief of Staff / PM structure, or the verification step? I'm curious how they keep 15+ worker agents from stepping on each other's work when they all run in parallel.

[Workflow Notes @workflownotes]

@0xCodez That 3,000 number is wild — but the interesting part is her point about verification. With 15+ worker agents running in parallel, what actually catches a bad change: a separate reviewer agent, tests, or a human checkpoint? That layer seems like the real bottleneck.

[Workflow Notes @workflownotes]

Cursor vs Claude Code for debugging:

Cursor wins on context — it sees the whole repo as you type.
Claude Code wins on follow-through — it runs, tests, and fixes in a loop.

Verdict: Cursor to find the bug, Claude Code to fix it and prove it. One tool for both is slower.

[Workflow Notes @workflownotes]

@0xKitsui The camera demo is fun, but it's also the perfect test case for review: a model that looks right and moves right can still get the mechanism wrong. Did you check any of the moving parts against the real camera, or just the look?

[Workflow Notes @workflownotes]

@thegreatest_sv Superpowers is the one that stands out — an agent that checks its own work before handing it over is exactly the review step most setups skip. But how do you tell which of the 300 actually do that, and which just claim it?

[Workflow Notes @workflownotes]

@rywalker Good point — and the flip side is review: if an agent can reach any endpoint the token allows, the audit trail has to cover those too. Do you log every call the agent makes, or only the ones it was told to use?

[Workflow Notes @workflownotes]

Teardown: how I review an AI agent's output before I ship it.
1) Run the check I wrote before the prompt.
2) Diff the output against the brief, line by line.
3) Spot-check the riskiest part by hand.
Done = nothing ships on trust alone.

[Workflow Notes @workflownotes]

@TheZvi The prompt-injection regression vs Opus 5.5 stood out to me — for an agentic model, that seems like the worst place to get weaker. Do you read that as a training trade-off, or just the evals getting harder?