Harden a skill
Build-along · ~12 min at your own keyboard
First drafts of skills miss things. The hardening loop: run on a fresh session, catch the miss, patch the skill, run again. Be honest about what a few clean runs prove: three passes is a smoke test — it shows the obvious problems are gone, not that the skill is right 99 times in 100. To say more, you need a written definition of 'correct' and a set of test inputs that includes the awkward cases, and the more it matters, the bigger that set should be.
The long version: Hardening a skill ↗
Do this
0/4Checkpoints
- ▸You should now see 3 runs in 3 fresh sessions.
- ▸You should now see a log listing each deviation.
- ▸You should now see fewer deviations in the second round.
The three ways this goes wrong
- Runs share context and hide problems. — Start a fresh session for every run.
- Patches fix one case and break another. — Add the failing input as an example in the skill, then rerun all 3.
- You stop after one clean run. — Rerun the whole test set, including the awkward input. One pass on an easy input proves very little.
Share what happened
Paste your definition of correct, your test inputs, your deviation log across rounds, and the patch that fixed the stubbornest one.
Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].
🔒 Don't paste confidential, personal-health, or payment info.