Optimize Prompt

Optimize Prompt

Optimize only from evidence. Prefer exact prompt checkpoints and digest metrics; use byte/token estimates only when exact tokenizer data is unavailable and label them as estimates.

Task

$ARGUMENTS

Phase 1: Load Evidence

Start with digest:

astra journal digest last --format json
astra journal digest <SESSION_ID> --format json
astra journal digest <SESSION_ID> --focus summary --format json

Then locate exact prompt payloads when needed:

ls -lt ~/.astra/sessions/<SESSION_ID>/step_checkpoints/*-heavy.json 2>/dev/null | head
ls -lt /tmp/debug-*-turn*-full.json 2>/dev/null | head

Evidence priority:

Do not optimize from vague impressions such as "prompt feels long".

Phase 2: Map Component Owner

Phase 3: Measure Waste

System:

• Identify repeated static sections, task-irrelevant rules, or conditional sections emitted without the matching capability/tool.
• Verify source before recommending removal; many sections are cache-stable and cheap after cache hits.

Tools:

• Compare visible_tools_count, tools_used_count, activated_tools_count, and actual tool calls.
• Waste signal: many visible tools plus low usage across repeated turns, or deferred tools activated but never called.
• Owner is tool surface/capability metadata, not ad hoc prompt text.

Skills:

• Check selected_skills and the actual user task.
• Waste signal: selected skill unrelated to the task or large skill instructions repeatedly injected.
• Fix by tightening trigger/description or deleting low-ROI skill content.

History/tool results:

• Inspect message sizes in heavy checkpoints.
• Waste signal: repeated file reads, huge tool outputs retained across turns, stale reasoning/tool results after compaction.

Budget:

• Use budget_pressure, context_ms, ttft_ms, compaction events, and turn token counts.
• Healthy sessions show pressure relief after compaction; sustained high pressure after compaction needs prompt/history/tool-result work.

Optional checkpoint size scan:

python3 - <<'PY'
import json, sys
path = sys.argv[1]
msgs = json.load(open(path, encoding="utf-8"))
for i, m in enumerate(msgs):
    role = m.get("role", "?")
    size = len(json.dumps(m, ensure_ascii=False))
    content = m.get("content", "")
    preview = content[:80].replace("\n", " ") if isinstance(content, str) else type(content).__name__
    print(f"{i:03d} {role:10s} {size:8d} bytes {preview}")
PY

Phase 4: Recommend Changes

Every recommendation needs:

• observed metric or checkpoint evidence;
• owning file/module;
• expected effect;
• verification command or digest metric to re-check.

Avoid:

• invented exact token savings;
• removing safety-critical instructions just because they are large;
• adding another prompt layer when the real issue is tool/skill selection metadata.

Output Contract

Observed:
- session=<id>, turns=<n>, pressure=<pattern>, visible_tools=<pattern>, selected_skills=<pattern>

Top savings:
1. <component> - <evidence> - owner=<file> - expected effect=<bounded estimate>
2. ...

Do not change:
- <large but necessary/cache-stable section, if any>

Verify:
- <digest/checkpoint/test command>

{
  "schema_version": 2,
  "skill": "optimize_prompt",
  "cause": "budget_pressure",
  "headline": "system prompt and tool surface contribute 60% of token budget with low tool utilization",
  "findings": ["visible_tools_count=45 but only 3 tools used across 12 turns"],
  "recommended_action": "defer rarely-used tools and trim system prompt static sections",
  "success_criteria": [
    {
      "metric": "budget_pressure",
      "operator": "lte",
      "threshold": 0.85,
      "window_turns": 3,
      "description": "sustained budget pressure drops below threshold"
    }
  ],
  "source": "real_skill"
}