HWAI Handoff

Use this as a thin wrapper over the project rule (see your repo rules). The handoff is operational state for the next agent, not a chat summary.

Quick Decision

• Same thread and short status update: answer normally; do not create a handoff.
• New session, another agent, cross-device work, context approaching limits, or a major milestone: create a handoff.
• Long logs/specs/URLs are inputs to compact first with context-prep; reference the artifact instead of pasting the body.
• Broad repo uncertainty: use retrieval first, then cite exact paths read.

Destination

Choose the narrowest durable target:

1. Temporary/private handoff by default: save outside the repo when the user only needs to paste into another session.
2. Repo handoff only when it is project state that future sessions should inherit.
3. Public-bound MCP/token-economy material: do not include private HWAI moat details. Apply the project rule (see your repo rules).

For repo handoffs, prefer a focused path such as notes/<topic>_handoff_<date>.md or an existing task/methods directory. Do not create HANDOFF.md in the repo root unless Greg asks for that exact file.

Format

Use the canonical # Compaction Handoff template from the project rule (see your repo rules). Always include these extra fields when useful:

• First action: the single next command/read/edit the fresh agent should do.
• Suggested skills: exact skills/rules to invoke, with paths.
• Files in progress: path plus state: editing, needs_review, partially_done, or ready.
• Public/private boundary: required when MCP stack, scraper, gateway, ContentOS, token-economy article, or public repo publishing is involved.

Quality Bar

A fresh agent with no chat history should be able to make real progress within 5 minutes.

Rules:

• Distill, do not copy conversation noise.
• Reference existing PRDs, plans, ADRs, commits, diffs, artifacts, and issue URLs by path/URL.
• Preserve exact user constraints, approval state, IDs, numbers, paths, dates, and blockers.
• Record what failed or should not be retried.
• Redact secrets, credentials, tokens, raw email bodies, private customer data, and PII.
• If refreshing an existing same-topic handoff, read it first, then rewrite from scratch; do not append stale context.

Evaluation

This skill is gated by skills/handoff/_evals/:

• trigger_activation.json checks the RU/EN handoff triggers and non-triggers.
• output_quality.json checks first action, suggested skills, files in progress, exact constraints, and public/private boundary.
• terminal_bench.mjs is the deterministic pre-score: good handoff artifacts must pass, summary-only or moat-leaking handoffs must fail.

Run:

node skill_eval.mjs --skill handoff --mock --no-trend-log
node skills/handoff/_evals/terminal_bench.mjs

Before promoting the skill for high-traffic use, run the real judge through the repo skill-eval gateway when available.

Claude Code With Large Context

A larger context window reduces handoff frequency; it does not remove the need. Create handoffs at clean boundaries and when prompt-cache stability, cross-session continuity, or multi-agent transfer matters.

For MCP token economy work, mention handoff as a cache-stability/context-discipline practice in articles, but keep tuned internal orchestration and measured moat details private unless Greg explicitly overrides the public boundary.