diff options
Diffstat (limited to 'assets')
| -rw-r--r-- | assets/codex-skills/fidget-spinner/SKILL.md | 9 | ||||
| -rw-r--r-- | assets/codex-skills/frontier-loop/SKILL.md | 4 |
2 files changed, 7 insertions, 6 deletions
diff --git a/assets/codex-skills/fidget-spinner/SKILL.md b/assets/codex-skills/fidget-spinner/SKILL.md index cfa3521..f187be3 100644 --- a/assets/codex-skills/fidget-spinner/SKILL.md +++ b/assets/codex-skills/fidget-spinner/SKILL.md @@ -52,7 +52,7 @@ If you need more context, pull it from: - `source.record` for imported source material, documentary context, or one substantial source digest; always pass `title`, `summary`, and `body`, and pass `tags` when the source belongs in a campaign/subsystem index - `note.quick` for atomic reusable takeaways, always with an explicit `tags` list plus `title`, `summary`, and `body`; use `[]` only when no registered tag applies - `hypothesis.record` before core-path work; every experiment must hang off exactly one hypothesis -- `experiment.open` once a hypothesis has a concrete base checkpoint and is ready to be tested +- `experiment.open` once a hypothesis has a concrete slice and is ready to be tested - `experiment.list` or `experiment.read` when resuming a session and you need to recover open experimental state - `metric.define` when a project-level metric key needs a canonical unit, objective, or human description - `run.dimension.define` when a new experiment slicer such as `scenario` or `duration_s` becomes query-worthy @@ -60,7 +60,7 @@ If you need more context, pull it from: - `metric.keys` before guessing which numeric signals are actually rankable; pass exact run-dimension filters when narrowing to one workload slice - `metric.best` when you need the best closed experiments by one numeric key; pass `order` for noncanonical payload fields and exact run-dimension filters when comparing one slice - `node.annotate` for scratch text that should stay off the main path -- `experiment.close` only for an already-open experiment and only when you have checkpoint, measured result, note, and verdict; attach `analysis` when the result needs explicit interpretation +- `experiment.close` only for an already-open experiment and only when you have measured result, note, and verdict; attach `analysis` when the result needs explicit interpretation - `node.archive` to hide stale detritus without deleting evidence - `node.create` only as a true escape hatch @@ -94,7 +94,8 @@ If the point is to test a claim, it should become a hypothesis plus an experimen Preserve one source digest if needed, then extract reusable claims into notes. 8. A hypothesis is not an experiment. Open the experiment explicitly; do not smuggle “planned work” into off-path prose. -9. Porcelain is the terse triage surface. Use `detail=full` only when concise +9. The ledger is scientific, not git-forensic. Do not treat commit hashes as experiment identity. +10. Porcelain is the terse triage surface. Use `detail=full` only when concise output stops being decision-sufficient. -10. When the task becomes a true indefinite optimization push, pair this skill +11. When the task becomes a true indefinite optimization push, pair this skill with `frontier-loop`. diff --git a/assets/codex-skills/frontier-loop/SKILL.md b/assets/codex-skills/frontier-loop/SKILL.md index e8f1b94..45235f4 100644 --- a/assets/codex-skills/frontier-loop/SKILL.md +++ b/assets/codex-skills/frontier-loop/SKILL.md @@ -58,7 +58,7 @@ DO NOT EMIT A FINAL TURN UNLESS YOU ARE ACTUALLY BLOCKED OR EXPLICITLY TOLD TO S ASSUME YOU ARE RUNNING OVERNIGHT. -1. Start from the current best checkpoint or most credible live branch. +1. Start from the current best accepted line or most credible live branch. 2. Study existing evidence from `fidget-spinner`. 3. Search outward if the local frontier looks exhausted or you are starting to take unambitious strides. 4. Form a strong, falsifiable hypothesis. @@ -68,7 +68,7 @@ ASSUME YOU ARE RUNNING OVERNIGHT. rerun only enough to understand the outcome. 8. Record the outcome through `fidget-spinner`. 9. Keep the line if it advances the objective or opens a genuinely strong new avenue. -10. If the line is dead, record that too, re-anchor to the best known checkpoint, +10. If the line is dead, record that too, re-anchor to the best known accepted line, and try a different attack. 11. Repeat. |