Appendix B Checklist Compendium¶

The book's seven checklists, print-ready. Sources at each list's end and in Appendix C.

B1 Minimum Viable Harness — side A (day one)¶

From zero to the first usable harness:

[ ] Write AGENTS.md/CLAUDE.md (project constitution: domain, stack, conventions, dangerous-operations list; ≤100 lines)
[ ] Write verification commands ("tests should run" → pre-completion check of results)
[ ] Create feature_list.json (features + pending/in_progress/completed)
[ ] Create a progress file (per-session progress record)
[ ] Set a completion gate (merges/PRs only in verified-green state)
[ ] Defaults: WIP=1; cut 80% of tools and add back on demand; start single-agent

Mindset: AGENTS.md is soft control ("please don't"); verification/permissions/hooks are hard control ("can't").

From "works" to "doesn't explode":

[ ] Habitual bottleneck trio: is the symptom at the Prompt layer (misreads), Context layer (forgets), or Harness layer (wrong execution path)?
[ ] An "error → rule" pipeline: every real incident becomes one rule/test/constraint
[ ] Run component shelf-life checks at every model upgrade (see B6)
[ ] Security baseline: pre-deploy (permission sandbox + danger list), in-execution (fail-closed interception), post-output (audit logs) — at least one item each
[ ] High-risk domains (ops/finance/healthcare): all five layers + human review node + rollback + change audit

For every loop intended to keep running:

[ ] Verifier gate independent of the generator (never trust the model's self-report)
[ ] Four stop paths written down: passed / achieved / over budget / rollback
[ ] Three-dimensional budget caps: tokens / time / cost — any one trips → stop
[ ] Rollback works (Git checkpoints) and the breaker path has been tested
[ ] State memory externalized (progress files / issues / Git); new sessions restore from files
[ ] Isolated workspace (worktree / temp branch)
[ ] Admission: recurrence × failure cost > threshold; order: deterministic API → single call → thin loop → heavy harness
[ ] Monitoring: success rate, regression rate, token efficiency; monthly review

On every model change or version bump:

[ ] Tool schema compatibility (names, parameter formats)
[ ] Citation/memory tag format consistency
[ ] System prompt structure (old hard rules may now be internalized — audit context responsibility ownership)
[ ] Transcript OOD risk (new model reasoning over old transcripts)
[ ] Prompt caches cleared; tokenizer changed (all caches invalid)
[ ] Benchmark: 20–50 real tasks, model axis × task axis, one variable at a time
[ ] Four signals: done right? process worse? old failures back? high-risk boundaries touched?

Trick: give sub-agents the other model rather than switching the main conversation's model.

#	Check	Pass criterion
1	Harness dry run done	Virtual requirement walked the full flow; gate/parameter/format bugs fixed
2	Quality gates programmatically checkable	"CI passes" decomposed into a decidable boolean
3	Maker and checker separated	Reviewer has caught maker-missed issues
4	Layered context loading live	Resident/triggered/on-demand layers; fill ≤ ~40%
5	Knowledge externalization started	Tacit constraints (field types/fallbacks/hot spots) in specs
6	Metrics updated	Delivery rate and review overhead tracked
7	External-call specs recorded	Timeout/retry/fallback in skills
8	Audit chain established	Full change directory per requirement (summary/tasks/review)
9	Knowledge lifecycle designed	draft/verified/proven + automatic decay
10	Team roles adjusted	Someone owns harness maintenance

Component	Encoded assumption	How to check
Context reset	Model has long-context anxiety	Off vs on, one round each
Sprint contracts	Model cannot self-decompose	Remove; does the generator wander?
Aggressive compression	Model distracted by redundancy	Loosen; does quality rise?
Tool formatting	Model can't handle raw formats	Raw vs formatted

Method: remove one component at a time. Build to Delete — a harness without a deletion plan is debt.

Any unanswered question: do not build yet.