| name | rocky-review |
| description | Wrap-up skill — runs rocky-checkpoint (extracts learnings from recent commits) followed by rocky-promptiq (rescores recent prompts), then prints a fused summary. Use as the default "end of work session" invocation. Trigger phrases: "/rocky-review", "rocky review", "wrap up the session", "end of session review".
|
Rocky · Review
Umbrella skill that runs the two end-of-session passes back-to-back so
the user only has to invoke one command. This is the recommended
default for "I'm done for now, capture what I learned and grade how I
asked for it." For CI / automation, prefer the underlying skills directly
— rocky-checkpoint after every push, rocky-promptiq on a slower
cadence (weekly cron, etc.).
The two underlying skills evaluate different things:
rocky-checkpoint is generative — it mines atomic durable concepts
from the commits and the conversation, and writes them to the PKG.
rocky-promptiq is critical — it scores the user's prompts against
a rubric for prompting quality.
Running them in sequence (rather than blended into one pass) is
deliberate. Switching evaluative stance is easier between two clean runs
than inside one — interleaving "what should I learn" with "how well did
I prompt" risks dragging one into the other (e.g. "this prompt was
vague AND so this topic is shallow," which is a category error).
Procedure
-
Run rocky-checkpoint to completion. Follow its skill file exactly
— read the queue, dedup against the global graph, extract topics,
add questions, add edges, drain the queue. The full skill body lives
at ~/.claude/skills/rocky-checkpoint/SKILL.md (or the OpenCode
equivalent); load it into your context if it isn't already, and
execute every step. Do not short-circuit it.
Capture the final report line so you can fuse it later:
Checkpointed N commits → M commit topics + P conversation topics, K merged (J across projects), Q questions added, E edges.
-
If checkpoint succeeded, run rocky-promptiq. Same rule — load and
execute the full skill at ~/.claude/skills/rocky-promptiq/SKILL.md.
It pulls unscored prompts, judges each against the rubric, persists
each evaluation via rocky prompt-eval, and prints a per-batch
summary.
Capture: number scored, batch average, lift over the heuristic
baseline, top patterns to work on.
-
If checkpoint failed (any non-recoverable error in step 1), do
not run promptiq. Report the checkpoint failure and stop.
Failure isolation is part of the design — promptiq is independent
work and there's no value in running it against a broken queue
state. The user can re-invoke /rocky-review after fixing the
underlying issue.
-
Print one fused summary to the user, format:
/rocky-review complete.
PKG · Checkpointed N commits → M new topics, K merged
(J cross-project), Q questions, E edges.
Prompts · Rescored N prompts. Avg <score>, lift +<N> vs heuristic.
Top pattern: <one-line synthesis>.
Two lines, two domains, no preamble. The user already knows what
they invoked.
-
Suggest rocky view for the live PKG graph + PromptIQ trend tab.
Rules
- Never blend the two passes. Running checkpoint and promptiq
concurrently in one LLM context tends to bleed evaluative stance
between them. Run them strictly sequentially with a clean handoff.
- The skills are the source of truth, not this file. If the
underlying skill files disagree with the fused summary format above,
follow each skill's own instructions and adjust the summary accordingly.
- Don't double-drain the queue. rocky-checkpoint runs
rocky checkpoint mark itself; this skill just invokes it.
- Don't double-score prompts. rocky-promptiq's
--unscored flag
already filters; trust it.
- CI integrations should bypass this skill and call
rocky-checkpoint or rocky-promptiq directly. Bundling two
skills makes sense for an interactive wrap-up, not for an automated
hook that should fail loudly on a single broken step.
Boundaries
This skill writes nothing of its own. All persistence is via the
underlying skills' rocky add-topic, rocky add-question,
rocky add-edge, rocky checkpoint mark, and rocky prompt-eval
calls. Same undo paths apply: rocky delete-topic <name> for any topic
recorded incorrectly; PromptIQ scores can be refreshed by re-invoking
the underlying skill.