make your agent review itself

Lenny's AI Builders - LAB 0015

one command reads a messy agent session and tells you what to fix, worst first.

my agent edited the last episode over 25 times.

so remember this: make your agent review itself.

why it matters

a long agent job leaves a long trail.

you know some parts felt slow. you just don't know which part to fix first.

Matt Pocock's new /retro skill reads that trail for you. it hands you one list, worst first.

what you get

one command you can run on any session that felt slow or messy.

• a ranked list. the worst problem sits at the top.
• real causes. it points at your setup, not at you.
• fixes you approve. nothing changes until you say yes.

the real example: the LAB 0014 edit

i pointed /retro at the session where my agent edited LAB 0014. there were 25 versions of that episode.

it found six things to fix. the top ones:

• 19 copied scripts. my Resolve fixes live in 19 copied scripts. one broke, so a version came out with no colour grade.
• i was the checker. i caught the overlap and cut-off pages myself. a check should catch them.
• 102 huge results. mostly full-size screenshots. the session filled up 4 times.
• raw footage by hand. getting it meant clicking through Descript. it should be one command.

my take: those are things i knew in the back of my mind. /retro just made them very clear.

try it

1. give your agent the link github.com/mattpocock/skills and say "install this repository of skills".
2. pick one session that felt slow or messy.
3. in that session, type /retro and hit enter.
4. read the list, worst first.
5. approve the fixes you agree with. fix one thing before your next run.

copy this prompt

run this inside the session you want reviewed. swap in your own job.

/retro
Look at this session - [the job that felt slow or messy].
What slowed you down, and what should I change in your setup so the next run goes faster? Worst first.

you'll know it worked when

you get one list, ranked worst first.

each item names what went wrong and what to change in your setup.

your next run of the same job is faster or cleaner.

a bad result: vague advice, items with no cause, or fixes you can't act on.

limits and dates

i ran /retro on 6 October 2026, on my own LAB 0014 edit session. your list depends on your session.

/retro suggests fixes. it doesn't prove they work.

the skill is in Matt Pocock's skills repo, version 1.3.1.

LAB 0015 went public on YouTube on 8 October 2026 at 6:40 am Perth time.

the rest of this LAB

• Claude gives you about 5x more. SemiAnalysis ran the same jobs on every big AI plan. on the $200 plans, Leo's screenshot of their data shows 28.6 billion tokens of Opus 5.5 vs 10.2 billion of GPT-6.1 Sol. 2.8x is raw tokens. 5x is SemiAnalysis's value at API prices.
• OpenAI's 28 days of gifts. Tibo says 28 days of gifts. day 1 is about 50% faster. Tibo's own posts - no changelog yet. my take: an advent calendar for people who just lost half their limits. Pro goes from 20x to 10x Plus.
• HyperFrames Studio. a video editor for agents. i haven't tried it.
• Beam. a giant open model you can't run. the weights aren't out and its scores are self-reported.
• ChatGPT has an app store now. you can submit a plugin to ChatGPT. 2,000% is MagicPath's own claim.
• only 4.5% of people pay for AI. a year ago it was 2.1%. a16z, US paid personal subscriptions, August 2026.

sources

• SemiAnalysis: Anthropic subscriptions offer 5x+ more value than OpenAI
• Leo (@synthwavedd): tokens by plan
• Tibo: 28 days
• Tibo: day 1, about 50% faster
• HyperFrames Studio
• Reflection AI: introducing Beam
• OpenAI: submitting a ChatGPT plugin
• Pietro Schirano: MagicPath growth
• a16z: top 100 gen AI consumer apps
• Matt Pocock's skills (v1.3.1, /retro)

next step

watch LAB 0015 on YouTube - the practical starts at 10:50.

want the next one? get the next field notes. optional - these notes stay free to read.