Stateful Skill Scaffold Generator

Overview

You are tasked with generating a new Stateful Agent Skill. While standard skills (via create-skill ) execute isolated tasks, stateful skills possess deeper systemic awareness: they manage artifact lifecycles over time, configure multi-dimensional tone, propagate epistemic confidence hierarchies, and link to other skills via Chained Commands.

These patterns were extracted from the L4 Anthropic Customer Support and Legal ecosystems.

Execution Steps

1. Requirements & L4 Pattern Discovery

Use a guided discovery interview. First, get the standard metadata (Skill Name, Description). Then, progressively ask the user which L4 State/Lifecycle templates they need injected:

Q1. Epistemic Trust (Tiered Authority) Does the agent need a Tiered Source Authority model to propagate a Confidence Score (High/Med/Low) into its outputs based on the evidentiary hierarchy?

Q2. Artifact Lifecycle Management Does this skill create or maintain persistent outputs (e.g., KB articles, tickets)? If so, we will inject the Artifact Lifecycle State Machine (Draft → Published → Needs Update) and a Scheduled Maintenance Cadence.

Q3. Multi-Dimensional Tone Configuration Does this skill draft external communications? If so, we will inject the Tone Configuration matrix (Situation Type × Audience Segment = Tone Label).

Q4. Escalation & Quality Gates Does this skill require an Escalation Trigger Taxonomy (Stop, Alert, Explain, Recommend) or a Business Impact Quantification Protocol before proceeding?

Q5. Workflow Navigation (Chained Commands) What commands logically follow this output? We will inject an "Offer Next Steps" block to chain this node to other skills.

Phase 1.5: Recap & Confirm

Do NOT immediately scaffold after the interview. You must pause and explicitly list out:

• The decided Skill Name and Description

• Which of the 5 L4 State/Lifecycle templates you plan to inject Ask the user: "Does this look right? (yes / adjust)"

1.8 Autoresearch Governance (Required)

Stateful skills often run iterative maintenance loops. Apply Karpathy-style governance:

• Baseline First: Record one baseline evaluation before modifications.

• One Variable Per Iteration: Change one control dimension at a time.

• Keep / Discard Decisions: Keep only measurable improvements; revert to last known good when performance regresses.

• Crash / Timeout Protocol: Treat failed iterations as explicit outcomes, log them, and continue from known-good state.

• Persistent Ledger: Write an iteration history to evals/results.tsv .

2. Scaffold the Infrastructure (Preventing Context Bloat)

Execute the deterministic scaffold.py script to generate the physical directories:

python3 ./scripts/scaffold.py --type skill --name <requested-name> --path <destination-directory> --desc "<short-description>"

3. Generate Lean Pattern References (Lazy-Loading)

CRITICAL: Do NOT bloat the generated skill with massive definitions of these patterns. Instead of writing out the entire theory of Escalation Taxonomies or Lifecycle State Machines in every new skill, you must practice Progressive Disclosure:

• For each selected L4 pattern in Step 1, create a LEAN file in references/ (e.g., references/tone-matrix.md ). Load its specific definition file from references/patterns/<pattern-name>.md to learn how to scaffold it.

• This file should ONLY contain the domain-specific tables (the actual matrix values for this specific skill).

• Do not explain how the pattern works; the central pattern-catalog.md already defines the mechanics. Just provide the blank or filled templates for this specific workflow.

4. Finalize the SKILL.md (Pointers Only)

Write the final SKILL.md . Ensure it:

• Keeps the primary instructions concise (<300 lines).

• Uses Markdown links (e.g., See Escalation Rules ) so the LLM only loads the context when needed.

• Includes the Chained Commands (Offer Next Steps) block at the bottom.

• Includes the mandatory Source Transparency Declaration.

• Includes an Iteration Governance section that references baseline-first + keep/discard + ledger behavior.

Next Actions

• Continuous Improvement Loop: Run ./scripts/benchmarking/run_loop.py --results-dir evals/experiments to evaluate your stateful skill with persistent experiment traces.

• Review Loop: Run ./scripts/eval-viewer/generate_review.py to launch the interactive viewer.

• Audit: Run audit-plugin to validate the generated artifacts.