Seven executable upgrades for rules, skills, state, tools, and verification—filtered to primary docs, maintained code, and reproducible methods.
The 2026 pattern is not more prompting. Keep prompts as maps, load procedures on demand, enforce invariants in code, persist replayable state, and let independent execution—not the actor’s confidence—decide “done.”
A minimal separation of routing, procedures, state, and hard gates.
AGENTS.md # map + universal commands
.agents/skills/verify-change/
SKILL.md # trigger + procedure
scripts/run.sh # deterministic gate
references/test-matrix.md # loaded only when needed
.agent/
runs.jsonl # append-only events
checkpoints/ # restart state
scripts/verify.sh # independent acceptanceEach item names an observable change you can make and test this week.
Keep topology, canonical commands, universal constraints, and “done” criteria in root AGENTS.md/CLAUDE.md. Put local rules beside the code they govern. Anthropic recommends under 200 lines; OpenAI reports roughly 100.13
One skill, one job. Put what + when in its description; move references and scripts out of SKILL.md. Invoke mandatory skills explicitly. Keep 10–20 positive, negative, and edge prompts; grade traces and artifacts—not vibes.24
Block destructive commands pre-tool, format/lint post-write, and run acceptance checks at stop. Unit-test command policy with match/not_match cases. Tool allow-lists and annotations are routing hints—not a security boundary.136
Persist actions, observations, failures, state deltas, and condensation events; derive the smaller prompt view from that log. Checkpoint run_id, status, attempt, last event, artifact URIs, and prompt/tool/serializer versions. Externalize bulk logs with reloadable pointers.7
Initialize {category, description, steps, passes:false}, a short progress file, and Git history. Each fresh session smoke-tests, completes one unfinished unit, verifies end-to-end, flips only its pass state, records failures, and commits cleanly.5
Bound searches; return file windows with paths, line numbers, totals, and omitted ranges. Include exit status and explicit empty-success. Make edits transactional: apply, syntax-check, roll back invalid mutations, show the changed region. Add tools only for repeated trace failures.8
BASELINE → simple loop first; ablate every new toolUse a fresh worktree or sandbox and an independent evaluator. The reproduction must fail before the patch, pass after, and regressions stay green. Record base commit, image digest, harness version, commands, outputs, and artifacts. Self-review is never the final oracle.159
DONE → executable evidence + clean state + durable handoff