Appendix DHow was the agent steered?
Chapter 04 says that BMAD prepares a brief for each story and that three reviewers read every change. This appendix looks at how that worked in practice: which model did what, how large a brief became, what the reviewers found and kept finding, and what only a person found.
Status, 2 November 2026
This appendix is a snapshot of work in progress, drawn from the project's records as they stood on 19 October. The findings were sorted from a sample, by Claude; only the totals of that sort were recorded, and a person's check of 20 of the 60 is pending. The improvements ranked at the end are proposals. Numbers in the form OPEN-nnn are entries in the project's register of open items, part of its quality system (chapter 06).
Which model did what?
Of the 172 stories, 142 record Claude Sonnet 5 as the developer, up to 24 September, and the 15 developed from 25 September record Claude Opus 5.5. The other 15, all from the configuration tool's first weeks, have no line for the model. From 25 September, the session logs that survive from that date show Opus 5.5 doing everything: the main session and the review agents it started. Before 25 September, the records do not say which model ran the reviews.
The switch was not a decision I recorded, and I cannot remember why I made it: it may have been for cost or for performance. It had a cost. In the six days after it, I hit the subscription's session limit seven times and its weekly limit once (chapter 01).
Did the switch make a difference? The records cannot say. The stories developed with Opus drew more review findings than those developed with Sonnet: about 15 per story, against about 9 per Sonnet story that records findings (75 of the 142). But they came later, they were larger and heavier on security, and the reviewers changed model at the same time. That is an observation, and nothing was measured.
How large did a brief become?
A story's brief, the story file before the developer's own notes, runs to a median of about 1,700 words, roughly 2,300 tokens, and to about 4,900 words at the largest. That is small beside what the model can hold. Larger briefs drew somewhat more findings, but the link is weak.
The context ran short in long sessions instead. The logs from late September to late October show the conversation compacted four times in a month. BMAD's answer is the one chapter 04 describes: a fresh, complete brief for each story, so that nothing depends on what a long session still remembers.
What did the reviewers find, and what does Claude repeat?
Nearly 900 findings are recorded. To see what kind they are, Claude read a fixed random sample of 60 and sorted them.
| Kind of finding | In the sample of 60 |
|---|---|
| Edge cases: an empty value, a label containing a separator, a person already inactive | 8 |
| Tests that are missing or assert too little | 8 |
| Comments or documents claiming what the code does not do | 8 |
| Display, styling and accessibility | 8 |
| Shared or stale state: a lost update, a list handed out for anyone to change, a resubmitted form | 7 |
| Security and audit | 4 |
| Error handling | 3 |
| Efficiency, hygiene and other logic | 4 |
| Dismissed, or a note | 10 |
Some kinds come back across stories. Documentation drifts from the code it describes; tests check less than they seem to; mutable state is shared where it should be copied; something is identified by its label where it should be identified by its key. The log of deferred findings flags one class of gap for the third time. These are the mistakes Claude repeats, and the reviewers catch them, one story at a time.
What did only a person find?
At least 27 stories record a defect found live, after the review had passed. Those I found myself include an option that did nothing when clicked, six pages nobody could reach, a patient form where "about everything is wrong", the Save of 17 September (Appendix B), a live update that had stopped working since an earlier release. The reviewers read the code. These defects showed only in a running browser, with real data, or when someone tried to use the product.
One shows the limit best. On 16 October, a page's draft was meant to lift a restriction on the Register action for Clerks. It also, by two clicks nobody intended, stopped Clerks from opening the patient page. That second change went through saving, approval under a second account and deployment, and every check held as designed, because none of them showed it. I found it while checking, live, what the deployed restrictions did. Chapter 09 tells what it led to: an approval protects only what the approver can see.
What did I learn?
The records show two things about Claude and BMAD. The reviewers find a great deal, nearly 900 findings, all but 18 of them in the 172 stories, most of them fixed in the same story, and they catch the same kinds of mistake every time they recur. And the defects that reached a running system were found by a person, using it.
About the method, each finding points at how the work was done:
| Finding | What it says about the method |
|---|---|
| The developer's model changed on 25 September, for a reason nobody recorded | The choice of model was never a decision, so it was never measured |
| Its effect on quality cannot be told from the records | Nothing compared the models on similar work |
| The same kinds of finding recur across stories | The reviewers correct each one, and nothing stops the next |
| 27 stories record a defect found live, after review | The review reads code; only a check on the running system sees what it does |
| The context ran short in long sessions | Sessions longer than one story lose what the brief does not carry |
What would I change?
These are proposals, ranked by the benefit I expect, each with the measure a later check would read.
| Rank | Improvement | Why at this rank | Measure |
|---|---|---|---|
| 1 | Every story ends with a recorded check on the running system, by a person or by an automated browser test | It targets the defects that passed review; it is also what OPEN-114 asks for | Defects found live after a story is done |
| 2 | The kinds of finding that recur become checks or instructions: a test-strength check (Appendix B), and a comparison of each comment that states behaviour with a test that proves it | The reviewers stop meeting the same mistake | Recurring findings per story |
| 3 | The model for each role is a recorded decision, and a change is tried on a few similar stories, measured, before it is adopted | It makes the choice explicit, with its reason, its cost and its effect | Stories per model, with their findings and their cost |
| 4 | One story per session | Nothing depends on a compacted conversation | Compactions per story |
Where does this leave the project?
At a gate. The agent was steered by a method that worked more than it failed: briefs small enough to hold, reviews that find a great deal, and a person who checks what runs. What it lacked was measurement: of the model, of the recurring mistakes, and of what each check catches. These four proposals are among the decisions waiting at the gate. A later version of this appendix will say which were taken.