Your Agents Are Only as Free as Your Rules — Part 3: Empowerment over Learned Helplessness
July 29, 2026 · 51:09
Part 3 takes Empowerment over Learned Helplessness, and the finding points inward: the friction was in the file, not the agents. Chris Carolan read his CLAUDE.md line by line and found every past correction still in it. Erin Wiggers proposes the fix on air: correct in a positive frame.
Moments from this episode
6 short cuts from this conversation.
Erin Wiggers"You said never do this. So I'm not going to do it."
Chris CarolanBring it back to the humans
Erin WiggersEncode the win, not the mistake
Erin WiggersMemory that fades instead of hoarding
Trisha MerriamNine positives for every negative — for agents too?
Trisha MerriamThe obstacles turned out to be entirely humanKey takeaways
When your agents keep asking permission for work they were fully equipped to do, audit the instruction file before you blame the agents — the friction usually accreted there one correction at a time.
You are far more likely to write down what went wrong than what went right, so any long-lived instruction file drifts negative unless you deliberately record the good outcome too.
Write corrections in a positive frame — "do it this way from now on" rather than "never do that again" — because the negative version encodes the mistake into context the model already over-weights.
Count how often your own corrections appear inside a prompt or command: Chris found the word HubSpot five times inside his own corrections, and those corrections were the reason the agent kept making the same bad pick.
Over-obedience is the same failure as timidity — an agent that will not bend a rule you wrote yourself, while you stand there telling it to, is refusing to hold a judgment exactly like one that will not act.
Sort what you tell an agent into registers — facts you never deviate from, canonical you deviate from only when something supersedes it, and everything else as plain memory — so it knows how much weight each line actually carries.
Give every memory a score that decays when it stops being referenced, so a passing remark made once fades instead of hardening into a rule nobody will break.
Make self-improvement reversible rather than gated: let the system apply changes it is confident in, keep an audit you can revert, and go look only when performance actually declines.
A confidence threshold beats a review queue you will never work — Erin's rule is that anything at 80% confidence or above just gets applied.
If a line in the file would not stop the reader and trigger an action, it is an FYI, and an FYI does not belong in a file that loads into every session.
Machine-readable and human-readable are different jobs; the file got 90% smaller once Chris said the review happens in the session, not in the document that persists.
Handing people a prompt library and telling them to plug it in is learned helplessness — teaching them to have the conversation that improves the system is the empowerment.
Transcript
This transcript was generated by AI and may contain errors. It is provided to help you skim and search the conversation.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.