Skip to content

Deliberate Currency

Ratified

Look again when a tripwire fires. Don’t look when none has.

  • Currency with AI tooling is an engineering practice, not spare-time reading
  • It has a named owner and a review cadence
  • Re-evaluation is triggered by tripwires agreed in advance
  • AI tooling decisions have a half-life of months
  • Nothing announces obsolescence; falling behind looks like normal
  • Scaffolding decays invisibly and may now make results worse
  • Churn is the mirror failure: re-litigating every release resets compounding investment
  • Tripwires fix both: we look when something meaningful changes, and only then
  • Inventory: every prompt, skill, and workaround, tagged with the model it was tuned against
  • Owner: a named person accountable for flagging invalidated skills
  • Sweep: on each major model release, ask whether each item still earns its keep
  • Kill test: does the skill still beat a plain prompt? Deletion is a win
  • External:
    • Frontier model release in a class we use daily
    • Pricing change beyond an agreed percentage
    • A pattern or protocol reaching real adoption among peers
  • Internal:
    • Agent retry rate climbing
    • Humans redoing agent output more often
    • Token spend per shipped change trending up
  • Each tripwire has two levels: “log it” and “re-evaluate now”
  • Time-box each re-evaluation so thresholds reflect its cost
  • Drift: skills tuned against a retired model still load, and nobody can say who owns them
  • Churn: tooling choices reopen on every release though no tripwire fired
  • Tripwires with no baseline: nobody can state retry rate or spend per shipped change, so nothing ever fires
  • An inventory that only grows: no sweep has deleted a skill
  • Can we state today’s cost per shipped change and first-pass success rate?
  • Owner by volunteer or rotation?

Proposals: Discussion #3, Discussion #4