
Capture every correction in a skill, so you never teach the same lesson twice.
Capture every correction you give an agent in a skill, so you never teach the same lesson twice. That one loop separates an AI that stays a toy from one that compounds, from a single skill up to a whole team.
A tree keeps every year: a wide ring for a good year, a narrow ring for a hard one.
Tanush Jagdish · June 2026
Capture every correction you give an agent, so that you never teach the same lesson twice. I repeat this in nearly every coaching session, because it separates an AI that stays a toy from one that compounds. Its mechanism is a loop: use a skill on real work, correct it, then fold the corrections back into the skill.
- 01
Use
Run the skill on a real task. A first pass gets you perhaps 40 to 50 percent of the way.
- 02
Iterate
Give feedback exactly as you would to a person. No borders. Smaller font. Three takeaways.
- 03
Enrich
“Take everything we learned in this chat and improve the skill.” The correction is kept for good.
- ↻ then back to Use, one rung higher
The first enrichment is the biggest jump. After a few cycles you stop doing the work and start coaching it. Percentages summarize coaching experience; they are not measurements.
#1From half right to nearly right
Run a skill on a real task for the first time, and it gets you perhaps 40 or 50 percent of the way. So you work with it, giving feedback the way you would to a person: smaller font, no borders, no em dashes, three key takeaways, action items by owner.
Then comes the step that matters. At the end of the session, tell the agent: take everything we learned in this chat and improve the skill, so I never have to say these things again. Your next run gets you to about 70 percent. Keep going and it reaches the 90s, and your job shifts from doing the work to coaching it.
Those percentages are a teaching illustration, not measurements. They summarize what I see across coaching clients: a skill usually needs four or five enrichment cycles to become consistently good, and within a week or two of regular use, most of the feedback you used to repeat is handled without prompting. Of all the cycles, the first produces the biggest jump, which is why I push people to do it at least once.
Research on language agents supports the mechanism. In Reflexion, a NeurIPS 2023 paper by Noah Shinn and colleagues, agents that wrote reflections on feedback and kept them in memory for later attempts improved markedly, reaching 91 percent on the HumanEval coding benchmark against a previous best of 80 percent. A skill file is a durable, human-curated version of that memory.
#2Do the work first, make the skill second
Order matters, and most people reverse it. Opening a chat to write the perfect skill from scratch produces the generic answer any model gives the first time, because it has not yet seen your taste or learned what you object to.
Instead, get to the output you want first: go back and forth, and give your ten points of feedback. Then package all of it into the skill. By the end of the session, the chat holds every correction you made. If you give ten points today, you will give the same ten tomorrow, and because the chat still holds them, the agent can write them into the skill.
In effect you are running an evaluation loop live, with a sample of one: draft, find what is wrong, fix it. When the draft looks good, often about halfway through the context window, ask the agent to review everything you worked on and update the skill. Two or three cycles usually suffice to start.
#3Why this loop matters most
Underneath the loop is one principle: no learning should die in a closed chat. Most people have taught a chatbot something real and watched it vanish when they closed the tab, only to teach it again the next day. That is the babysitting trap: “don’t do this, do that,” forever. Say something once and it should be recorded in a skill and reused. My blunter version for workshops: if you say something twice, it needs to be a skill.
A client a month into coaching described the stakes in his own words. He had noticed that each new session starts without the previous session’s context, so after fixing each problem he asked the agent which lessons belonged in the skill:
“If I don’t fold in those learnings, I’m in trouble.”
a coaching client, one month into the program
A skill is where corrections go to live instead of die.
#4A long game
Set expectations honestly: the best skills do not become great in a week. My own slide skill, which turns an idea into a finished PDF deck, took more than a year of iteration and still has room to improve; each fix took its own day of wrestling and one more rule folded in. When clients three weeks in are frustrated that a skill is not perfect, I tell them the truth: they are trying to do months of work in three weeks.
Duration is also the payoff. As the curve climbs, you stop being the loop. Early on, you check everything. After long use of a well-enriched skill, you may barely check its output, not from laziness but because it has been reliably right. Approving every step stops making sense for that kind of work, though not for high-stakes work, where Inflection Points argues for deliberate checks. An enrichment curve is also a trust curve.
#5Not a bet against the models
A common worry: why spend a year enriching a skill if the next model makes it obsolete? Rich Sutton’s 2019 essay “The Bitter Lesson” is the usual reference. Across 70 years of AI research, he argues, general methods that scale with computation have beaten approaches built on hand-coded human knowledge. Read quickly, that argues against investing in skills.
We read it the other way. A skill does not replace the model’s reasoning; it records what good output looks like for your specific work, which no general model can know in advance. When a stronger model reads the same skill, it applies those standards better. So set up your data and skills so that your agents improve automatically when the models do. Enrichment is how you aim each model improvement at your own work.
#6From one skill to a brain to a team
This pipeline scales in two directions.
Up into your personal brain, covered in Part 4 of our curriculum: the loop runs on your CLAUDE.md file as well. Notice a mistake, correct it, add a line, and the mistake does not recur. A personal brain is the enrichment pipeline applied to everything, not just to one skill.
Out into the team, covered in Part 6: skills deepen rather than multiply. When a venture capital client sent us 27 feature requests for its agent, we triaged them: 18 extended skills the agent already had, 4 called for new capabilities, and 5 were already covered. That is one client, but the lesson travels: deepen the skills you have before building new ones.
Extend
Deepen a skill the agent already has.
New
A genuinely new capability to build.
Covered
Already handled today.
Skills deepen in layers. We don’t need 27 new skills; we need to deepen the ones we have.
On a managed agent, enrichment becomes the roadmap. For the same client, the plan was to grow the deal-intake skill in layers, from parsing pitch decks and writing a basic summary toward founder profiles, competitive landscapes, network maps, and a first-draft investment memo, each layer resting on the one beneath it.
Every correction you capture also adds to something larger: a record of how your firm works, specific to you, that improves with use.
#7Keep going
- Where the loop starts: Part 3 of our curriculum, Skills.
- Same loop, applied to your instructions and memory: Part 4, on the personal brain.
- Same loop at team scale: Part 6, on the institutional brain.
- Why a rising curve lets you relax enforcement: the research note Inflection Points.
- How the skill layer shows up across professions: the research note Convergence.
No chat should be wasted, and no learning should die in a closed chat. If you say something twice, make it a skill. Every correction you fold in compounds, from one skill to your brain to your team.
More from our research.
Working on the same question?
We look for people to interview, data to compare against, and teams willing to test what we find. Leave a note and we will write back.