Project Ops Protocol (POP)

Proven on a long-running multi-session project: 100+ docs, 50+ tools, several parallel AI sessions and one human decision-maker — without losing state, duplicating work, or colliding. Everything below is copy-paste-able into any new project. It is a method, not a company secret.

When to use

• You are starting a real project that will span many sessions/agents/days.
• Several AI sessions (or people) work the same project in parallel.
• Someone asks "how do we organize / what's the mechanism / do a retrospective".
• Skip it for one-off questions and single-tool jobs.

Core idea: four layers

Plan     layered battle map: Stand → Build → Distribute; each layer closes
         before the next opens; an honest "watchlist" for what we deliberately
         do NOT do now (with a review cadence).
Execute  measure first, reuse before build, converge over duplicate;
         statuses change only after they actually happen.
Sync     one append-only alignment board + milestone heartbeats + @channel tags.
Memory   one relay note (read first by any new session) + a living handbook
         where process changes land BEFORE they are announced.

Standard artifacts (build once, reuse always)

Hard rules (the cards that prevent the common failure modes)

1. Heartbeat: within 5 minutes of finishing a deliverable, +1 line on the alignment board (time | line | did what | path/url). No polling loops; to request something from another line, tag it on the board.
2. No-collision: before starting work, read the division of labour + the tail of the alignment board. Reuse before rebuild: scan existing assets first.
3. Unique anchors: when editing shared files, match on a unique string and re-check neighbouring lines after saving (prevents silent overwrites).
4. Measure first: plan conclusions come from actual inventory (counts/APIs), never from vibes.
5. Honest watchlist: what we are NOT doing goes on the watchlist with a review date — it never becomes a zombie todo.
6. Secrets: a local encrypted credential store only; secrets never in chat, commands, screenshots, or shared plaintext folders. Inject into processes via environment, echo nothing.
7. Gate before external publish: desensitize scan → quality judge (5 dims, ≥0.80 pass, any 0.0 vetoes) → safe packaging → owner approval.
8. Process changes land in the handbook before being announced anywhere else.

Quality judge (LLM-as-judge rubric)

Score each deliverable 0.0–1.0 per dimension; pass ≥0.80; any 0.0 = veto:

1. Factual accuracy — every claim traces to a source?
2. Citation accuracy — sources actually support the claims?
3. Completeness — conclusions, boundaries, attribution, license all present?
4. Source quality — primary/official preferred over secondary retells?
5. Discipline — no internal names, secrets, local paths, or private strategy words in anything going outside? Quick 30-second self-check: (a) can each claim be pointed to a source? (b) am I citing the primary source, not a retell? (c) would this leak anything internal if it went public?

New-project checklist (day 1)

• battle map v0.1 (layers + milestones + watchlist)
• three ledgers identified/created
• alignment board created + heartbeat rule announced (multi-session only)
• existing assets scanned (no rebuilding what exists)
• roles/boundaries written down where everyone can see

Wrap-up checklist (project or big day ends)

• layer close-out checked, including watchlist explanation
• retrospective written (data / mechanism / traps / take-aways)
• relay note updated (single place every new session reads first)
• open items listed explicitly (≠ failure — honesty is the standard)

Boundaries & philosophy

• Markdown + filesystem is enough until retrieval actually degrades; do not front-load vector DBs / knowledge graphs / event buses (benchmarks show plain files beat specialised memory tools when retrieval is reliable).
• Handbook-first (borrowed from GitLab), orchestrator-worker + compressed reports back (borrowed from Anthropic multi-agent research), judge-gated publishing. All ideas are credited to their sources — copy what works.

© 2026 Zhao Xinghua (Steven Zhao) — CC BY 4.0 · MedXpert × SynomosAI