Meta Loop Efficiency

/meta-loop-efficiency

Weekly audit of startup-loop skill files. Read-only scan. Writes one committed artifact per run.

Invocation

/meta-loop-efficiency                     # default scan, commit artifact
/meta-loop-efficiency --threshold <N>    # custom line threshold (default 200)
/meta-loop-efficiency --dry-run          # print findings; do not commit
# implementation path: pnpm --filter scripts startup-loop:meta-loop-efficiency-audit -- --dry-run

Operating Mode

AUDIT ONLY — no skill file modifications permitted. Writes and commits one artifact per run (skipped on --dry-run). Use the script-backed audit engine in scripts/src/startup-loop/diagnostics/meta-loop-efficiency-audit.ts for report generation.

Audit Scope

In-scope: all lp-* prefix skills + startup-loop + draft-outreach. Out-of-scope: idea-*, meta-*, ops-*, review-*, biz-*, guide-* — never scan these. Canonical list derived from the startup-loop Stage Model in startup-loop/SKILL.md.

No scope override is supported by design. Non-loop skill families have different invocation patterns and their own improvement cadences. For a full cross-portfolio scan, run this skill against each family directory manually as a one-off operator action.

Heuristics

All heuristics operate over ALL .md files in each skill directory (SKILL.md + modules/ + others), not SKILL.md alone.

H0 — Duplicate candidate detection

SHA256-hash the content of each in-scope SKILL.md (normalize whitespace). Group skills with identical hashes. Emit a "Possible duplicates" section for any group of ≥2. Advisory only; does not affect H1–H3 scoring.

H1 — Size / modularization

Threshold applies to SKILL.md orchestrator line count only:

• wc -l SKILL.md > threshold AND no modules/ → monolith (strong signal)
• wc -l SKILL.md > threshold AND has modules/ → bloated-orchestrator (weaker signal)
• wc -l SKILL.md ≤ threshold → compliant

Additional: if modules/ exists and any module exceeds 400 lines → module-monolith advisory (logged, not ranked).

H2 — Dispatch adoption gap

• dispatch_refs_any_md: count of subagent-dispatch-contract matches across ALL .md files in the skill directory
• phase_matches_any_md: matches of anchored heading pattern ^#{1,6}\s+(Phase|Stage|Domain|Step)\s+[0-9]+ across ALL .md files
• dispatch_refs_any_md == 0 AND phase_matches_any_md ≥ 3 → dispatch-candidate

Implementation note (H2): Use grep -h to suppress filenames before counting — grep -c with a single file omits the filename: prefix, causing awk -F: to produce wrong counts. Correct pattern:

find "$skill" -name "*.md" -exec grep -hE "^#{1,6} (Phase|Stage|Domain|Step) [0-9]+" {} \; | wc -l

H3 — Wave dispatch adoption

Applies only to skills that reference lp-do-build as their execution skill AND have phase_matches_any_md ≥ 3 (filters out skills that merely name lp-do-build as a downstream step). If wave-dispatch-protocol.md is not referenced in any .md file in the directory → wave-candidate (advisory).

H4 — Deterministic extraction opportunity

Detect skills where deterministic content likely belongs in executable/data artifacts (.ts, .json, .yaml) rather than prose:

• deterministic_signal_count: count matches across all skill .md for algorithmic markers (regex, schema, validation contract, sort, dedupe, threshold, parse, normalize, map, filter, step-by-step decision table)
• artifact_ref_count: count references to .ts, .json, .yaml, .yml in the skill directory docs
• If deterministic_signal_count ≥ 6 AND artifact_ref_count == 0 → deterministic-extraction-candidate

Advisory intent: trigger an idea to move repeatable logic into typed code/data artifacts with tests, not just split markdown.

H5 — Shrink-without-simplification (anti-gaming)

Flag suspicious "line target passes" that likely just reshuffle prose:

• Compare current vs previous audit per skill using:
  • SKILL.md line delta
  • total .md footprint delta in the skill directory (SKILL.md + modules/ + other .md)
• If SKILL.md decreased by at least 20 lines but total .md footprint increased, and no new .ts/.json/.yaml/.yml references were introduced → shrink-without-simplification

This is advisory and should be called out as a quality risk, not counted as success.

Ranking

Produce three separate hierarchical ranked lists. Do not merge H1, H2/H3, and H4/H5 signals.

List 1 — Modularization opportunities (H1)

Rank within each tier by SKILL.md lines descending:

1. high tier — monolith or bloated-orchestrator
2. medium tier
3. low tier

List 2 — Dispatch opportunities (H2 + H3)

Rank within each tier by phase_matches_any_md descending:

1. high tier — dispatch-candidate or wave-candidate
2. medium tier
3. low tier

List 3 — Deterministic extraction and anti-gaming (H4 + H5)

Rank by severity:

1. critical — H5 shrink-without-simplification
2. high — H4 deterministic-extraction-candidate on high-tier skills
3. medium — H4 on medium-tier skills
4. low — H4 on low-tier skills

Invocation tier assignments (fixed)

• high: startup-loop, lp-do-build, lp-do-plan, lp-do-replan, lp-do-sequence, lp-offer, lp-channels, lp-seo, lp-forecast, lp-do-fact-find
• medium: lp-launch-qa, lp-design-qa, lp-experiment, lp-design-spec, lp-prioritize, lp-site-upgrade
• low: lp-onboarding-audit, lp-assessment-bootstrap, lp-readiness, lp-baseline-merge, lp-measure, draft-outreach

Any lp-* skill not listed above defaults to low tier. Add new skills here when they enter the loop.

Delta-Aware Planning Anchor

Always write the artifact. Emit planning anchor ONLY when findings are new:

1. Locate previous artifact: glob docs/business-os/platform-capability/skill-efficiency-audit-[0-9][0-9][0-9][0-9]-[0-9][0-9]-[0-9][0-9]-[0-9][0-9][0-9][0-9].md, sort lexicographically, take last. Files not matching this exact pattern are excluded (lexicographic order equals timestamp order for this format).
2. For each skill flagged in List 1, List 2, or List 3:
   • Not in previous artifact → new-to-HIGH → emit planning anchor.
   • Already flagged → status: known (suppress anchor for that skill).
3. Any previously compliant skill now flagged → always emit planning anchor (regression).
4. No new items and no regressions → emit:

   "No new HIGH opportunities since last audit (YYYY-MM-DD-HHMM). Known opportunities remain open — see previous anchor."

Output Artifact

Path: docs/business-os/platform-capability/skill-efficiency-audit-YYYY-MM-DD-HHMM.md

Required header (first 6 lines of artifact body):

scan_timestamp: YYYY-MM-DD HH:MM
threshold: <N> lines
scope: lp-*, startup-loop, draft-outreach
git_sha: <HEAD SHA 7 chars>
previous_artifact: <filename or "none">
skills_scanned: <N>

Artifact sections:

1. Scan summary — header + scanned/compliant/opportunity counts
2. Possible duplicates (H0)
3. List 1 — Modularization opportunities (H1, hierarchical)
4. List 2 — Dispatch opportunities (H2/H3, hierarchical)
5. List 3 — Deterministic extraction + anti-gaming (H4/H5, hierarchical)
6. Planning anchor — when new-to-HIGH items exist: suggest /lp-do-fact-find startup-loop-token-efficiency-v2
7. Delta status — new-vs-known breakdown, regression flags

Commit Guard

Before writing/committing the artifact:

1. Run git status --porcelain.
2. Stage ONLY the artifact file by explicit path: git add docs/business-os/platform-capability/skill-efficiency-audit-<stamp>.md
3. Verify with git diff --cached --name-only that only the artifact is staged.
4. If any unrelated files appear staged → abort; warn operator; do not commit.
5. Use writer lock: scripts/agents/with-writer-lock.sh --wait-forever -- git commit

Note: a dirty working tree (other unstaged changes) does NOT block the commit, provided only the artifact file is staged. Use --dry-run to inspect output without committing.

Heuristic Evolution

To extend or update this skill:

1. Add new heuristic H<N> section under ## Heuristics.
2. Update ## Ranking lists to include the new signal.
3. Add tier entries for any new lp-* skills added since last update.
4. Keep SKILL.md ≤200 lines; extract to modules/heuristics.md if needed.