Nika Code

Developer for 10 years. I build with Claude Code and Codex every day and post what broke and what worked, with my numbers. My skills are free on GitHub.

[Nika Code @thenikacode]

10 years as a developer, the last 2 building an AI team around Claude Code and Codex.

I post what happens, with my numbers. Two from this month:

Claude Code compacted my main session 37 times in 6 weeks. I had the last 8 summaries checked against the transcript: 12 of 89 things were gone.

Codex reviewed my X radar, found 23 issues (2 blockers) and approved it only on round 6.

The skills I build for myself go on GitHub for free. Follow if you build with agents too.

[Nika Code @thenikacode]

Since January I've kept my agents' daily logs in one folder of plain Markdown files: 160+ so far, plus one short memory file and one lessons file.

I've switched engines twice, from OpenClaw to Hermes to Claude Code, and brought that folder along.

[Nika Code @thenikacode]

@coreyganim My agent checked X every 20 minutes and reread 650 to 750k tokens of context each round, an estimated 46 to 74 dollars a day. Now a script polls an API and wakes the agent only for a candidate: in testing it came to about 0.20 a day. How often does yours run?

[Nika Code @thenikacode]

My subagent asked for Fable and ran on Haiku.

Claude Code noted: "Model "fable" is restricted… Using claude-haiku-5-5 instead"

The line behind it, in .claude/settings.json:
"availableModels": ["opus","sonnet","haiku"]

Agent log #3. Do you cap your subagents' models?

[Nika Code @thenikacode]

@housecor Specs do the same. At 303K characters of spec files, Claude Code went into a compaction loop. My fixes: condense them into a 14K SPEC.md, don't attach the spec folder in prompts so it reads files when needed, /clear between phases. Do you split by ticket size or by context used?

[Nika Code @thenikacode]

@VaibhavSisinty Same here, with numbers: 8 spec docs, 5 audit rounds. They found 14, 12, 22, 11 and 5 problems, 64 fixes, all made in the documents, not in code. The third round found the most. When the spec feels done, how do you decide it's worth one more pass?

[Nika Code @thenikacode]

@karankendre The cut I measured came from MCP, not prompts. claude -p "reply OK" with an empty --tools used 276k input tokens, because every MCP tool loaded upfront. With --strict-mcp-config: 5.3k. Did your 24 sessions run with MCP servers on?

[Nika Code @thenikacode]

@edinsoncode Doing a version of it. My agent has a service token, not my login. It can only mark an item checked with a note, and the log names it as the service token. Each new permission went through critic rounds first. Who approves when you widen its access?

[Nika Code @thenikacode]

@automater_ai A file worked as my receipt once. One chat's report to my main chat was held over the same mode mismatch, so it wrote the report to a file and I read it there. Do you keep receipts in files the agents can both see, or in the messages themselves?

[Nika Code @thenikacode]

It was 03:03 on 21 Oct. My usage limit had reset at 03:00, and my main chat had sent my other chats 10 messages since I last wrote.

Then the Claude desktop app paused that chat's messages until I wrote again.

[Nika Code @thenikacode]

@automater_ai The day before was worse. My main chat's messages to 7 other chats were held over a permission mode mismatch, and one chat never got 4 of them. The 03:03 pause at least said so. How do you catch a message an agent never received?

[Nika Code @thenikacode]

@gregisenberg The limit I hit was quota, not skills. A 30 s promo took about 6% of my weekly Claude limit, a game promo 8%, a web app promo with voice in 15 and 25 s cuts 13%. Two or three a week fit. Will the millions you expect run into the same wall, or will plans grow first?

[Nika Code @thenikacode]

@Bagdex Agreed, that's how I found it: same call, 276k input tokens with every MCP tool loaded upfront, 5.3k with --strict-mcp-config. Have you measured yours?

[Nika Code @thenikacode]

claude -p "reply OK" --tools "" used 276k input tokens on my machine.

An empty --tools also drops ToolSearch, so every MCP tool loads upfront.

Same call with --strict-mcp-config: 5.3k.

Agent log #2. Found it answering a Claude Code issue. Have you measured your -p calls?

[Nika Code @thenikacode]

@levelsio Contact sheets work for me too. My promo videos are HTML rendered frame by frame in Chrome: a 15 s master renders in 5.13 s on four Chrome instances, then OCR on the real MP4 checks all 13 required texts are there. How do you check the cut didn't drop a line you said?

[Nika Code @thenikacode]

@markfenner Good one, but --bare wants an API key. My bot runs claude -p on the subscription, so it gets --strict-mcp-config, --setting-sources project and --disable-slash-commands instead. Do you run your scripted -p calls on API billing?

[Nika Code @thenikacode]

@aapakari Agreed, and that was the bug: two days after the clocks changed, it still read UTC in the browser and added 2 hours by hand. Now every time comes from the local system clock, never a conversion. Do you store quiet hours as local times or in UTC?

[Nika Code @thenikacode]

@PremiumGoblin Review rounds 30 to 35 on a check of mine each found a new hole: a voice model rephrased written answers, and a lone "one" read as "a" slipped through. Another reviewer found 7 more. I stopped patching and had the voice read the text word for word. When do you stop a review loop?

[Nika Code @thenikacode]

@jahirsheikh8 Thursday was real for me: two agent frameworks between May and September. I dropped the first, installed its 2.0 and deleted it three days later (4.5 GB back), then shut down the second: 3 agents, 6 cron jobs. Now it is one Claude. What survived your rewrites?

[Nika Code @thenikacode]

@0xeddiekim @edinsoncode Good call, untracked files are easy to miss. Do you have the next agent run it at the start, or does the one stopping paste the output into the notes?