The seam that broke the system
- Observed
- Producer and consumer disagree
- Root closed by
- Contract test + caller trace
- Residual
- Unobserved paths remain explicit
Supporting work · Reusable AI workflows
Make the work inspectable before asking someone else to trust it.
Agent Skills Lab packages three difficult AI workflows into guided intake, concrete artifacts, deterministic checks, and an explicit human decision boundary.
Designed, built, and evaluated by Titus Lai.
Current statePUBLIC REPOSITORY · LIMITED EVIDENCE
Evidence workbench
Three jobs. Three artifacts someone else can open.
Each artifact keeps the judgment, evidence, unresolved risk, and required human decision in one place.
INPUT→INSPECTED→SEALED
Two judgments kept apart
- Decision lens
- Useful enough to act on?
- Source lens
- Does the evidence support it?
- Gate
- Material source failure can block
The shot before generation
- Continuity
- Reference roles declared
- Timing
- Action and camera beats compiled
- Spend
- Generation not auto-authorized
Three Skills
What each Skill changes in the work
01
agent-system-integration-auditTrace the whole path before naming the defect.
A plausible fix is dangerous when nobody has followed the real producer, adapter, and consumer path.
- It delivers
- A neutral detection record, an evidence-bound audit packet, and the remaining unobserved risk.
- Human boundary
- Read-only by default. Finding a defect does not grant repair authority.
Professional packet quality 0/3 → 2/3Semantic detection stayed 3/3 → 3/3
02
dual-lens-investment-reviewKeep a persuasive report from outrunning its sources.
Decision usefulness and source integrity are different judgments; mixing them hides material failure.
- It delivers
- Two isolated reviews, frozen typed delivery, then deterministic reconciliation.
- Human boundary
- A material source-integrity failure can veto publication. It never gives investment advice.
Valid lens delivery 0/6 → 6/6Reconciliation 0/3 → 3/3 · raw reasoning not scored
03
ai-anime-production-directorMake the shot inspectable before generation.
A beautiful prompt cannot resolve contradictory continuity, unclear references, or missing approval.
- It delivers
- A timed storyboard, reference roles, a human approval receipt, and a compiled shot contract.
- Human boundary
- Conflicts stop at a human gate. The Skill never silently authorizes generation or spend.
Valid storyboard/workflow artifacts 0/2 → 2/2Blind preference unavailable · finished-media quality unproven
Bounded evaluation
Better structured delivery in this set—not a universal intelligence claim.
The result is an offline quality regrade of preserved outputs from a frozen synthetic paired set.
- Source
- 25 preserved result receipts
- New spend
- Zero new model calls
- Coverage
- One paired run per case
- Claims gate
- EXPAND_BEFORE_PUBLIC_CLAIMS
Measured deliveryWithoutWithBounded reading
Audit packetWithout0/3With2/3
Packet quality improved; detection did not.
Dual-lens deliveryWithout0/6With6/6
Typed delivery improved; reasoning lift was not scored.
Shot workflowWithout0/2With2/2
Preproduction artifacts improved; finished media was not judged.
Every case has one paired run. This is useful engineering evidence, not a statistically strong public benchmark.
Claim boundary
What this work does not prove
- 01General intelligence or cross-model capability uplift
- 02Broad defect-detection lift
- 03Raw reasoning lift for the dual-lens review
- 04Investment performance or trading outcomes
- 05Finished-media quality or blind preference
- 06Production certification
Release state
The repository is public. The evidence boundary stays explicit.
The curated package is now available on GitHub after independent review, fresh-clone verification, and secret scanning.
- Repository
- Boss0320/agent-skills-lab
- Rights
- All Rights Reserved · portfolio evaluation only
- License
- Not open source; reuse, modification, and redistribution are reserved
- Public state
- Verified 31 August 2026 · reviewed main a74026e
Inspect the public repository↗Opens GitHub in a new tab.