Pulled one step out of my Level Scout agent into a standalone command: /alarm-check. It checks that every level I've marked actually has an alert on it. Useful on its own — I chart when I chart, not at 08:00. Reads live alerts and price from TradingView, flags dead alerts and bad labels. Read-only. It reports, it never deletes. Two things came out of it. Building the skill exposed a gap in my agent spec — my label vocabulary had no slot for one of my setups. Added it. The skill only counts what's on the chart, and that's the point. I only mark levels where I'd take a trade. Everything else is noise. One rule: never report "this level has no alert" — only "no alert found via TradingView." Absence of a check is not a passed check. Next: the command that writes the agent's state file. Right now my two skills can't talk to each other.

Agent runs on two machines, so I wrote two system prompts. Then saw it: identity and rules now lived in two files that could disagree. Same drift that bit me two weeks ago. Collapsed it to one file with a lookup at the top. Then hit a question I couldn't answer — what actually triggers the second machine? My agent and I were holding two different versions of the architecture. Neither of us had gone back and read the files. An hour on a problem that didn't exist. Fifth time I've had that conversation with it. "Read the files first" was already written in three places. Saying it doesn't make it stick — so today it went into the file the agent loads every run, as a rule with the failure attached: an answer from memory looks exactly like an answer from the source. No alarm goes off. Three things I'm changing on my side, because half of this was mine: - New chat every day. My project loads its context on session start. Carrying yesterday's window forward means that never fires. - Say the goal out loud first. Without it, the agent can't tell a detour I want from one it should pull me out of. - Never let it guess what we're doing. If it doesn't know, ask or look it up — don't fill the gap. Your agent will misremember right alongside you. Make it quote the source, not summarise it.

Audit said zero new MCPs needed — everything was already connected. Then I checked what the tools actually expose. The MMT server has 15 tools. None of them do alerts. My spec had a step telling the agent to create alerts there. It could never have run. Read the underlying script too: it computes its own levels and can't accept mine. Spec and script disagreed about who owns the levels, and nobody had noticed. Fix: flipped the step from create to verify. The agent now checks every candidate level has an alert and names the gaps. Better job anyway — it catches the alert I forgot. Also had to write in what it can't see: it reads TradingView alerts, not MMT ones. So it reports "no TradingView alert found", never "no alert". Absence of a check isn't a passed check.

Translated nine days of SOP into instructions the agent actually runs. Two things worth passing on. I designed a labelling convention. It was wrong. My agent reads price levels I mark on charts. So I specified a schema — FIB_0.618, RB_VAH, CH_POC — so it could parse them reliably. Then I labelled some levels the way I actually think: "0.618 - for longs", "TP1 POC of range - move SL to entry". Mine were better. They carry the role — what to do at that level — not just what the level is. And the agent is an LLM. It reads natural language fine. I'd built a rigid schema to solve a parsing problem that doesn't exist. New rule: write how you think. Only the role needs to be recognisable, because that's what decides what the agent does with it.

Read my own SOP back as if I'd never met myself. Found 24 gaps in a document I'd spent six days writing. Two patterns did most of the damage. Contradictions from my own edits. Day 36 tightened a threshold. Day 37 changed a rule. Day 38 moved a deadline. Every time the later section was right — and the earlier one still said the old thing. Six of those. An agent reading top to bottom hits the stale version first. Steps that never existed. My outputs section described a screenshot. My error handling described a sync failure. Neither had a step that produced them. Six missing steps, invisible until I traced backwards. The check that caught them: for every output and every failure mode, ask which numbered step produces this? If you can't point at one, it doesn't happen. One more, and it stung: two tools reported success on things that hadn't happened — including my own symbol search returning 15 results, which I read as "doesn't exist." It did exist. A negative finding needs the same verification as a positive one.