Idea from a colleague's own skill design: keep a gotchas.md file alongside
a skill's SKILL.md, instructed to be read only when the skill actually
fails or misbehaves on a concrete setup — not on every invocation. It
accumulates specific, hard-won fixes ("this failed on X because Y, the fix
was Z") rather than living inline in the main skill body, where it would
bloat every read for a problem most invocations never hit.
This is a genuinely different failure mode than what SKILL.md's existing
"Gotchas" sections cover: those are known-upfront caveats, always loaded. A
gotchas.md would be for surprises discovered in the field, on someone's
specific machine/environment, that don't belong in the always-loaded body
but are worth not re-discovering from scratch next time. Anyone hitting a
new one appends to it.
Candidate first skills to try this on: setup-ci (already has a
## Gotchas section that's grown organically from real bootstrap sessions —
a natural test of whether splitting stable vs. field-discovered gotchas is
worth it) and nix-binary-cache (machine-specific Nix problems are exactly
the kind of thing this is for).
Idea from a colleague's own skill design: keep a
gotchas.mdfile alongsidea skill's
SKILL.md, instructed to be read only when the skill actuallyfails or misbehaves on a concrete setup — not on every invocation. It
accumulates specific, hard-won fixes ("this failed on X because Y, the fix
was Z") rather than living inline in the main skill body, where it would
bloat every read for a problem most invocations never hit.
This is a genuinely different failure mode than what
SKILL.md's existing"Gotchas" sections cover: those are known-upfront caveats, always loaded. A
gotchas.mdwould be for surprises discovered in the field, on someone'sspecific machine/environment, that don't belong in the always-loaded body
but are worth not re-discovering from scratch next time. Anyone hitting a
new one appends to it.
Candidate first skills to try this on:
setup-ci(already has a## Gotchassection that's grown organically from real bootstrap sessions —a natural test of whether splitting stable vs. field-discovered gotchas is
worth it) and
nix-binary-cache(machine-specific Nix problems are exactlythe kind of thing this is for).