Content
81%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is a highly actionable, well-sequenced routing guide with excellent validation and dry-run discipline, backed by real, well-signaled reference files. Its main cost is length: repeated base-URL and guardrail warnings plus a few concept restatements inflate token usage, and several domains could move to references to slim the overview.
Suggestions
State the base-URL rule (never hard-code api.letta.com, keep LETTA_BASE_URL paired with the runtime's LETTA_API_KEY) once in a single prominent place and reference it from the REST-helper and local-settings sections instead of repeating it three times.
Trim restatements of concepts Claude already knows, e.g. "Permissions decide whether tool calls are allowed, denied, or require approval" and "Compaction controls how old messages are summarized when context is evicted", down to the product-specific deltas.
Move the account-usage/billing and channels sections into new one-level-deep reference files (e.g. references/usage-and-billing.md, references/channels.md) and keep short routing summaries in SKILL.md, reducing overview token load while preserving discoverability.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most content is dense and product-specific (CLI flags, backend quirks) that Claude could not know, but the base-URL/hard-coded-Cloud warning recurs in at least three sections and a few passages restate known concepts ("Permissions decide whether tool calls are allowed, denied, or require approval"), so it could be tightened more than anchor 4's "minor instances". | 3 / 5 |
Actionability | Nearly every section gives copy-paste-ready commands with real flags — `letta model set --model-settings '{"max_tokens":16384}'`, full `npx tsx … --dry-run` invocations, a permission-rule JSON example, and `openssl rand -hex 32 | letta secret set WEBHOOK_TOKEN --stdin` — covering the common cases for each layer. | 5 / 5 |
Workflow Clarity | The "Safe workflow" is an explicit six-step checklist (scope, inspect + rollback patch, dry run, smallest change, verify, report) with validation after every write, `--confirm-*` gates on destructive operations, and error-recovery loops for broken state (out-of-band recovery, stale-binary diagnosis), matching the anchor-5 pattern. | 5 / 5 |
Progressive Disclosure | Three one-level-deep reference files (model-settings, compaction-prompt-patterns, api-patch-examples) are all present, clearly signaled inline ("Read … before changing reasoning") and in a References section, plus a scripts table; the gap keeping it below anchor 5 is that the ~440-line body inlines several whole domains (usage/billing, channels, CLI flags) that could be split into references. | 4 / 5 |
Total | 17 / 20 Passed |