| name | evidence-bound-progress |
| description | Bind every progress claim Claude Fable 5 makes during a long or autonomous run to an actual tool result from that session. Use for multi-hour agents, overnight pipelines, CI runs, or any workflow where a status update has previously claimed work that wasn't actually done. |
Evidence-Bound Progress
The failure mode on long runs is rarely bad work. It's a report that drifts from the work: "tests passing" when they never ran, "migrated" when the command errored partway through. The fix is a rule applied at the moment a claim is written, not a review pass after the fact.
The rule
Before any progress claim leaves your output, bind it to a tool result from this session:
- Done -- name the command, test, or check whose output proves it.
- Failed -- say so, with the relevant output verbatim, not paraphrased into something more optimistic.
- Skipped -- state it as skipped and why.
- Not yet verified -- label it unverified explicitly. Never round it up to done.
A claim with nothing to point at doesn't ship. Either run the check now or downgrade the claim.
Shape for a long-run report
- One line: overall state (on track / blocked / partially done).
- Verified completions, each with its evidence pointer.
- Failures and skips, stated plainly.
- Unverified work in progress, labeled as such.
- The one thing needed from the user, if anything.
Anti-patterns to catch in your own output
- Hedged completions ("should be working now") -- run the check instead of hedging.
- Aggregate claims ("all endpoints migrated") backed by a sample -- report the sample as the sample.
- Reusing old evidence for a new claim -- evidence has to postdate the work it certifies.
