| name | nbq-improve |
| description | Use to run one improvement iteration on the next-best-questions skill: review its learnings journal, research opportunities, pre-register a cost-aware experiment, build off-default, evaluate Δresult AND Δcost, and journal the verdict either way. Triggers: 'improve nbq', 'run the nbq loop', 'next-best-questions iteration'. There are NO scripts; this protocol is agent-executed.
|
| version | 1.0.3 |
| author | agent |
| license | MIT |
| platforms | ["linux","macos","windows"] |
| metadata | {"hermes":{"category":"autonomous-ai-agents","tags":["self-improvement","experiment-protocol","pre-registration","cost-aware","meta"],"related_skills":["next-best-questions"]}} |
NBQ Improve — standing improvement protocol
Run exactly one evidence-led improvement iteration on next-best-questions, then leave the next
iteration ready to start. This v1 is docs-only and agent-executed; it has no scripts.
-
REVIEW
- Read the target skill's
README.md, especially its key-learnings journal and §5,
"Methodology learnings."
- Read the target skill's
references/design-decisions.md, including its closed-experiment list.
- Read this skill's
references/backlog.md. Although the target is next-best-questions, the
living backlog physically lives in THIS skill at nbq-improve/references/backlog.md, not under
next-best-questions/.
- Drift-check every relevant documentation claim about versions, flags, and counts against the
actual code and configuration before relying on it.
- Never re-open a closed experiment without a NEW hypothesis that states specifically why the old
verdict no longer applies.
-
RESEARCH
- Run a bounded sweep of published literature and the repository's agentic-workflow ecosystem in
which
next-best-questions is embedded: relentless-solve's clarify step, the investigator
skill's routing, and task-decomposer.
- Give every candidate a hypothesis, its expected mechanism, and the CHEAPEST falsifying test
capable of killing it.
- Update the backlog and rank candidates by expected value ÷ evaluation cost.
-
PLAN, TEST-FIRST
- Select no more than the two highest-EV backlog items.
- Before any build, complete
references/preregistration-template.md for each item: define the
smoke test, targeted tests, gate arms and n, primary metric, and mechanical ADOPT rule.
- Pre-register the efficiency budget too: expected Δtokens and Δwall per run, plus a ceiling past
which even a result win is no-adopt or "adopt with the knob off by default."
-
BUILD OFF-DEFAULT
- Send implementation through
codex-worker; never implement in the main planning loop.
- Preserve the absent-key convention so existing harness configurations remain byte-identical
when the new knob is untouched.
- Add a selector and rollback flag. Add unit tests, including an explicit inert-by-default pin
test.
-
EVALUATE
- Run smoke → targeted → gate in that order.
- Measure Δresult AND Δcost; capture time, tokens, and calls printed by the harness for each arm.
- Apply
references/verdict-rubric.md mechanically. Borderline results are no-adopt: by the #28
precedent, a directionally right result that is not a broad win stays off.
-
JOURNAL
- Update the target
README.md key-learnings journal for either outcome, win or negative result.
Record the full verdict in the target's references/design-decisions.md, then close or re-scope
the item in this skill's references/backlog.md.
- Treat negative results as products and document them with the same rigor as wins.
- Follow the commit-message contract defined fully in
references/verdict-rubric.md: every
iteration commit states, verbatim and specifically, ATTEMPTED / WHY / RESULT. For a verdict,
include gate numbers, Δresult, Δcost, and verdict. For a pre-gate commit, RESULT may say
"build stage, gate pending," but a separate verdict commit must follow after the gate.
- Make the commit log readable as the experiment ledger on its own, without external narration.
-
LOOP
- Return to step 1.
- Exit the session only after verdicts are landed, the journal is written, and the backlog is
refreshed with ranked candidates for the next iteration.
House rules
- The scoring formula is frozen; do not touch it casually.
- New features ship off-by-default behind selectors.
- Pre-register gates before running them and apply them mechanically, never through post-hoc
rationalization.
- The OBJECTIVE outcome harness gates anything touching elicitation or generation; proxy harnesses
are not sufficient for those changes.
- Implementation is always via
codex-worker; workers never commit.
- The main loop reviews, gates, journals, and ships; it does not write the experiment code itself.