Quality Gates.
Five gates for evaluating a completed agent package. Run before delivering. Every blocker must pass. No more than two majors remaining. The 02-golden-example/ folder is the reference standard.
Pass Criteria
Run these gates before handing a package to the human. The quality file itself enforces the discipline — if a blocker fails, the package is not shippable.
All blockers pass, majors ≤ 2
Every blocker passes across all five gates and no more than two majors remain. Package is shippable as-is.
Blockers pass but 3+ majors remain
Flag for human review. Name the majors explicitly in the handoff summary so the human can decide whether to ship or iterate.
Any blocker fails
Fix before delivering. Do not put the package in front of the human with an open blocker.
Weights: Gate 1 (25%), Gate 2 (25%), Gate 3 (20%), Gate 4 (15%), Gate 5 (15%).
Gate 1 — Completeness (25%)
Does the package have all five files with all required sections?
All five files exist
01-start-here.html with all three parts (A, B, C). 03-process.html with numbered, sequential steps. 04-context.html with an input manifest. 02-golden-example.html (or the 02-golden-example/ folder for meta-agents). 05-quality.html with at least three gates.
Hard fork on golden example mode was asked
The build asked the human, before any mapping or generation, whether they were providing the golden example (Mode A) or the builder was drafting it (Mode B), and the human’s answer is recorded plainly in the session log (e.g. “Proceeding in Mode A” or “Proceeding in Mode B”). A package whose build skipped the fork — or silently defaulted into a mode the human did not choose — is not shippable.
The new agent carries its Why
The generated agent’s process file opens with a 00 — Why is this important? section, before step one. It holds a real, laddered reason that connects the deliverable up to a business outcome (deliverable → near outcome → business result → OGSM) — not blank, not a restated task, not generic filler like “to grow the business.” If the Why is missing or fluff, the package is not shippable.
Required sections present
START-HERE Part A includes what it produces, what it needs, estimated time, and how to run. Process has clear phase boundaries and decision rules for branching. Context categorizes inputs as required, recommended, or optional. Golden example is a real, completed deliverable — not a placeholder. Step 02b and Step 02c are present in 03-process.html Phase 1.
Nice to haves
Files cross-link to each other via nav. Time estimates are realistic (not placeholder “TBD”).
Gate 2 — Accuracy (25%)
Is the content correct, internally consistent, and grounded in the source material?
Factual correctness
Process steps match what the agent actually needs to do (not hallucinated). Context file inputs match what the process file references. Golden example is consistent with the process — following the steps would produce something like the example.
Internal consistency
File paths referenced in START-HERE match actual file names. Clarifying questions are specific to this agent (not generic boilerplate). Quality gates are specific to this agent (not generic boilerplate).
Precision
Time estimates account for each phase. Domain-specific terminology is used correctly.
Gate 3 — Format (20%)
Does the package follow the canonical house style and structural conventions?
House-style compliance
All files are valid HTML (not markdown, not plain text). All internal files embed the canonical <style> block from 0. Chief of Staff/house-style.html verbatim. Only Inter is loaded as a font — no Instrument Serif, no DM Sans, no JetBrains Mono, no Helvetica.
Structural conventions
START-HERE uses the three-part structure (Manifest, Agent Training, Human Training). Process uses numbered steps with phase groupings. Context uses the .rule-group/.rule-card pattern with Required/Recommended/Optional eyebrows. Golden example uses the deliverable’s format — not the agent format or the house style.
Polish
Sticky nav with logo left, links to sibling files right. Auto-numbered sections render (the accent-red 01, 02… appear above each h2). Footer has the red dot. Responsive at 768px breakpoint.
Gate 4 — Executability (15%)
Could a fresh Claude session read the START-HERE and successfully run this agent?
Self-contained execution
Process file is complete enough that a new session could follow it without additional context. All external file paths referenced in the process actually exist (or are clearly marked as “human provides”).
Operational clarity
Decision rules cover the common edge cases — not just the happy path. Required human inputs are explicitly listed so the human knows what to provide. Output format is specified: what files to create, where to save them.
Resilience
Graceful handling when optional inputs are missing. Suggested next steps after first run.
Gate 5 — Value (15%)
Is this agent worth building? Does it save meaningful time and improve with use?
Justification
The agent automates work that has been done at least twice before, or is clearly going to recur.
Compounding potential
Agent improves over time as context files accumulate. The three-run graduation path is plausible (Run 1 → Run 2 → schedulable). The agent’s level is appropriate and a level-up path is noted.
System integration
Agent fits cleanly into the GAAPC folder structure. Trigger phrase is natural and memorable. Agent could be registered in the Chief of Staff agent directory.