optimize

Files (2)

• SKILL.md
• optimizer-loop.md · references/

SKILL.md (first lines)

# Optimize

<run>
This skill is a standing prompt. The campaign is in
`${CLAUDE_SKILL_DIR}/references/optimizer-loop.md`. Read `common.md`, then the loop file, and execute it
as written: it is the instruction set for this run, not background.
</run>

<paths>
- Plugin root: `${CLAUDE_PLUGIN_ROOT}`
- This session's id: `${CLAUDE_SESSION_ID}`
- Shared rules: `${CLAUDE_PLUGIN_ROOT}/references/common.md`. Read it before the loop file.
- Loop-state script: `python3 ${CLAUDE_PLUGIN_ROOT}/scripts/loop_state.py`
- AGENTS.md template: `${CLAUDE_PLUGIN_ROOT}/templates/AGENTS.md`
</paths>

<parameters>
Start from the campaign's `<parameters>` block and apply these overrides from
the invocation: $ARGUMENTS

Record the resolved block at the top of the run record with the baseline. An
override that names no parameter goes under "Needs from me" rather than being
guessed into one.
</parameters>

<prerequisites>
- The campaign cites sections of the repo's AGENTS.md by name, above all
  Measurement (real path, primary benchmarks, instruments, hazards). If they
  are missing, the template is at the path above;
  propose the fill under "Needs from me" rather than writing it yourself.
- If the primary benchmarks are unnamed, or their guardrails and
  counter-checks are, recommend `/design-scoreboard` first. A speed target with
  no quality guardrail is a number the campaign can reach by breaking things.
</prerequisites>

<after_compaction>
Re-read the campaign file and the run record before the next action. The
campaign's instructions do not survive compaction on their own; the record
does.
</after_compaction>

From fblissjr/fb-claude-skills, Apache-2.0.

More from fblissjr/fb-claude-skills