Alpha Batcher's banner

Alpha Batcher

@alphabatcher • 51,859 subscribers

I love AI

Shorts

I thought my AI agent needed a bigger context window wrong the leak was my shell every time it ran git diff HEAD~5 the model ate 92,000 tokens after ztk 18,000 ls -la src 2,000 tokens became 53 cargo test passing 397 tokens became 21 the fix is a 260KB Zig binary that sits between your agent and the shell it compresses command output before it hits the model not by summarizing everything by knowing what each command actually needs to preserve > diffs keep changed lines > tests keep failures > ls keeps structure > logs dedupe repeats > JSON and errors pass through untouched 5.8M tokens saved in one 256 command session stop buying bigger context windows while feeding your agent raw terminal sewage

I thought my AI agent needed a bigger context window wrong the leak was my shell every time it ran git diff HEAD~5 the model ate 92,000 tokens after ztk 18,000 ls -la src 2,000 tokens became 53 cargo test passing 397 tokens became 21 the fix is a 260KB Zig binary that sits between your agent and the shell it compresses command output before it hits the model not by summarizing everything by knowing what each command actually needs to preserve > diffs keep changed lines > tests keep failures > ls keeps structure > logs dedupe repeats > JSON and errors pass through untouched 5.8M tokens saved in one 256 command session stop buying bigger context windows while feeding your agent raw terminal sewage

17,516 Aufrufe

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

Boris Cherny: "My job is to write loops." 5-day version you can copy for Claude Code Day 1: repo memory - write CLAUDE.md - include stack, test commands, code style, release rules, gotchas - add .claude/settings.json with the shell commands Claude can run Day 2: verification skill - pick 1 flow Claude keeps breaking - put the browser/API test in .claude/skills/ /scripts/ - make it return: pass/fail, failing step, screenshot/log path Day 3: commands Add these files: .claude/commands/babysit.md - checks your PRs - reads CI - handles obvious review nits - surfaces design questions .claude/commands/triage-issues.md - labels new issues - dedupes against existing ones - assigns owners .claude/commands/deploy-watch.md - checks the live app - reports regressions - avoids touching production Day 4: loops > /loop 5m /babysit > /loop 15m /triage-issues > /loop 5m /deploy-watch Day 5: overnight work - schedule /morning-report - schedule /deep-audit - write results to .claude/inbox/ - let your morning loop read from that folder The rule: every code-writing loop gets a separate verifier. Builder makes the change. Verifier runs the real app. You read the diff. Skip that and you wake up to 14 broken PRs with very confident summaries. Video: "Reflecting on a year of Claude Code" with Boris Cherny & Cat Wu

Boris Cherny: "My job is to write loops." 5-day version you can copy for Claude Code Day 1: repo memory - write CLAUDE.md - include stack, test commands, code style, release rules, gotchas - add .claude/settings.json with the shell commands Claude can run Day 2: verification skill - pick 1 flow Claude keeps breaking - put the browser/API test in .claude/skills/ /scripts/ - make it return: pass/fail, failing step, screenshot/log path Day 3: commands Add these files: .claude/commands/babysit.md - checks your PRs - reads CI - handles obvious review nits - surfaces design questions .claude/commands/triage-issues.md - labels new issues - dedupes against existing ones - assigns owners .claude/commands/deploy-watch.md - checks the live app - reports regressions - avoids touching production Day 4: loops > /loop 5m /babysit > /loop 15m /triage-issues > /loop 5m /deploy-watch Day 5: overnight work - schedule /morning-report - schedule /deep-audit - write results to .claude/inbox/ - let your morning loop read from that folder The rule: every code-writing loop gets a separate verifier. Builder makes the change. Verifier runs the real app. You read the diff. Skip that and you wake up to 14 broken PRs with very confident summaries. Video: "Reflecting on a year of Claude Code" with Boris Cherny & Cat Wu

141,110 Aufrufe • vor 2 Monaten

> built the app with Claude > shipped to App Store > told X it was live > got downloads > opened the dashboard > nothing > checked the code > StoreKit looked fine > checked App Store Connect > Paid Apps Agreement not active > banking missing > tax forms unfinished > Small Business Program never applied for > Apple still taking 30% > first payout still 45 days after fiscal month close > external payment copy sitting in one forgotten settings screen > reviewer would have found it instantly The app was not broken, the business setup was before you launch an iOS app, do this: > set up banking > complete tax forms > sign Paid Apps Agreement > apply for Small Business Program if under $1M > create real product IDs > test sandbox purchase > test restore > test cancellation > test subscription grace period > remove every sloppy external payment reference > then submit vibe coding gets you the app App Store Connect gets you paid

> built the app with Claude > shipped to App Store > told X it was live > got downloads > opened the dashboard > nothing > checked the code > StoreKit looked fine > checked App Store Connect > Paid Apps Agreement not active > banking missing > tax forms unfinished > Small Business Program never applied for > Apple still taking 30% > first payout still 45 days after fiscal month close > external payment copy sitting in one forgotten settings screen > reviewer would have found it instantly The app was not broken, the business setup was before you launch an iOS app, do this: > set up banking > complete tax forms > sign Paid Apps Agreement > apply for Small Business Program if under $1M > create real product IDs > test sandbox purchase > test restore > test cancellation > test subscription grace period > remove every sloppy external payment reference > then submit vibe coding gets you the app App Store Connect gets you paid

175,107 Aufrufe • vor 3 Monaten

Karpathy compressed the agent problem into 1 sentence: "If you can't evaluate then you can't auto research it, right?" That's the rule I keep coming back to with long-running coding agents Before you launch /goal or /loop, write the verifier: - what counts as done - what evidence proves it - which checks run every pass - which artifact gets saved - which failure sends it back into the loop Then let the agent run The loop can keep going because proof sits outside the agent's own explanation Tests, screenshots, benchmark curves, browser runs, changed files That's how you get autonomy without babysitting a transcript for 6 hours Read the full breakdown on goals, verifiers, loops, artifacts, and session memory in the article below

Karpathy compressed the agent problem into 1 sentence: "If you can't evaluate then you can't auto research it, right?" That's the rule I keep coming back to with long-running coding agents Before you launch /goal or /loop, write the verifier: - what counts as done - what evidence proves it - which checks run every pass - which artifact gets saved - which failure sends it back into the loop Then let the agent run The loop can keep going because proof sits outside the agent's own explanation Tests, screenshots, benchmark curves, browser runs, changed files That's how you get autonomy without babysitting a transcript for 6 hours Read the full breakdown on goals, verifiers, loops, artifacts, and session memory in the article below

43,806 Aufrufe • vor 2 Monaten

> installed Hermes Agent > launched several skills > opened outputs folder > empty > checked the config > MEMORY_PATH was on ./data/memory.db > restarted the process > the folder was created anew > all records are gone > the memory was reset with each restart > the scheduler was triggered at 3 a.m. > timezone was UTC > skills were not picked up without restarting > SKILLS_WATCH has been disabled > failed skill just dropped silently > SKILL_RETRY is disabled by default > outputs lived separately from obsidian > i had to carry it with my hands Hermes was not broken. The config was default. before you run Hermes as a 24/7 operation, do this: > set an absolute path for MEMORY_PATH > specify the correct SCHEDULER_TIMEZONE > turn on SKILLS_WATCH=true > configure CONTEXT_FILES for preload > increase MEMORY_RETRIEVAL_DEPTH to 20 > send OUTPUT_PATH to Obsidian vault > enable Telegram notifications > setting up chaining between skills > add nightly memory consolidator > enable SKILL_RETRY with alerts

> installed Hermes Agent > launched several skills > opened outputs folder > empty > checked the config > MEMORY_PATH was on ./data/memory.db > restarted the process > the folder was created anew > all records are gone > the memory was reset with each restart > the scheduler was triggered at 3 a.m. > timezone was UTC > skills were not picked up without restarting > SKILLS_WATCH has been disabled > failed skill just dropped silently > SKILL_RETRY is disabled by default > outputs lived separately from obsidian > i had to carry it with my hands Hermes was not broken. The config was default. before you run Hermes as a 24/7 operation, do this: > set an absolute path for MEMORY_PATH > specify the correct SCHEDULER_TIMEZONE > turn on SKILLS_WATCH=true > configure CONTEXT_FILES for preload > increase MEMORY_RETRIEVAL_DEPTH to 20 > send OUTPUT_PATH to Obsidian vault > enable Telegram notifications > setting up chaining between skills > add nightly memory consolidator > enable SKILL_RETRY with alerts

45,793 Aufrufe • vor 2 Monaten

540,000 lines of Rails is a brutal way to learn the agent era Garry's List shipped with: > 262,000 lines of app code > 276,000 lines of tests > 127 background jobs > 33 cron jobs > a 1,778-line claim verifier Then GStack hit 105k GitHub stars Do not copy the Rails pattern 1. Finish the task with an agent 2. Say `skillify it` 3. Save the markdown skill 4. Keep TypeScript only for deterministic I/O 5. Add a unit test for the code 6. Add an LLM eval for the skill 7. Add an integration test across both 8. Add a resolver so the agent loads the skill next time A Rails feature serves one app A skill pack can run against the next hackathon, repo, customer inbox, research folder, or QA run Audit your repo this week Cut code that only exists to: > re-check model output > retry the same model call > validate what the model already has context for > schedule work because the agent has no reusable procedure > preserve a one-off prompt in chat history Keep the code that touches files, APIs, auth, billing, devices, databases Move the judgment into markdown Test the markdown Reuse the skill

540,000 lines of Rails is a brutal way to learn the agent era Garry's List shipped with: > 262,000 lines of app code > 276,000 lines of tests > 127 background jobs > 33 cron jobs > a 1,778-line claim verifier Then GStack hit 105k GitHub stars Do not copy the Rails pattern 1. Finish the task with an agent 2. Say `skillify it` 3. Save the markdown skill 4. Keep TypeScript only for deterministic I/O 5. Add a unit test for the code 6. Add an LLM eval for the skill 7. Add an integration test across both 8. Add a resolver so the agent loads the skill next time A Rails feature serves one app A skill pack can run against the next hackathon, repo, customer inbox, research folder, or QA run Audit your repo this week Cut code that only exists to: > re-check model output > retry the same model call > validate what the model already has context for > schedule work because the agent has no reusable procedure > preserve a one-off prompt in chat history Keep the code that touches files, APIs, auth, billing, devices, databases Move the judgment into markdown Test the markdown Reuse the skill

35,693 Aufrufe • vor 2 Monaten

Prompt engineering is where people go when they refuse to write files Karpathy's LLM Wiki gist is painfully simple: > raw sources stay untouched > the agent writes the wiki > `CLAUDE.md` or `AGENTS.md` tells it the rules > `index.md` maps the pages > `log.md` records what changed That is why a good answer should not die in chat history If the model gives you a useful comparison, summary, decision, bug explanation, or research thread, file it back into the project Use: > `wiki/topics/[topic].md` > `wiki/entities/[person-or-company].md` > `wiki/sources/[source-name].md` > `wiki/analysis/[question].md` Make the agent keep the wiki from rotting: > backlinks > summaries > contradictions > stale claims > orphan pages > missing source links For coding projects, same idea: > `CLAUDE.md`: who you are, project rules, files not to touch > `index.md`: current files and features > `log.md`: date, change, reason > saved pages: decisions you do not want to re-explain Start by saving one good answer today

Prompt engineering is where people go when they refuse to write files Karpathy's LLM Wiki gist is painfully simple: > raw sources stay untouched > the agent writes the wiki > `CLAUDE.md` or `AGENTS.md` tells it the rules > `index.md` maps the pages > `log.md` records what changed That is why a good answer should not die in chat history If the model gives you a useful comparison, summary, decision, bug explanation, or research thread, file it back into the project Use: > `wiki/topics/[topic].md` > `wiki/entities/[person-or-company].md` > `wiki/sources/[source-name].md` > `wiki/analysis/[question].md` Make the agent keep the wiki from rotting: > backlinks > summaries > contradictions > stale claims > orphan pages > missing source links For coding projects, same idea: > `CLAUDE.md`: who you are, project rules, files not to touch > `index.md`: current files and features > `log.md`: date, change, reason > saved pages: decisions you do not want to re-explain Start by saving one good answer today

14,041 Aufrufe • vor 2 Monaten

Keine weiteren Inhalte verfügbar