Verax Vault

Canva templates, AI workflows & digital tools for architects, designers & studios. Practice kits:

[Verax Vault @VeraxVault]

1/
16 open-source GitHub repos that do the work of a video editor, a web agency and a dev team.

most people scroll past these. don't.

YouTube videos, websites, agent harnesses, all inside 🧵

(bookmark it, #11 alone is worth it)

[Verax Vault @VeraxVault]

@qumarl_hoda Very doable, but plan in months, not weeks. Order that works: Python and APIs, one LLM app end to end, RAG with a small eval set, then a simple tool-calling agent. Ship 2–3 with a README on how you measured quality. Free interview Q&A sample if useful: https://t.co/HvWM2DwLbY

[Verax Vault @VeraxVault]

@Dagimlogs Start from real traces. Hand-label 30–50 runs, turn each failure into a check (right tool, right args, outcome verified in the target system), and rerun that set on every prompt or model change. "Done" should be checked against system state, not the agent's own report.

[Verax Vault @VeraxVault]

Don't let the model own this. The tool layer derives an idempotency key per logical refund (say, from order_id), sends it on every retry, and the payment API dedupes. Stripe's Idempotency-Key header works this way. On timeout, check the order's refund status, then retry with the same key.

[Verax Vault @VeraxVault]

@JonathanMONDAUT One option: keep a canonical AGENTS.md in a small shared repo and have each repo pull it in (a submodule or a tiny sync script in CI or pre-commit), with repo-specific rules in a short local section below. Then a quarterly visit starts with one sync command.

[Verax Vault @VeraxVault]

@halfchargedd Turn on tracing that logs every tool call with its inputs and raw outputs (LangSmith, Langfuse, or OpenTelemetry-style spans all do this). Then have the agent cite the source IDs it used, so you can check its claims against what it actually retrieved.

[Verax Vault @VeraxVault]

india's coastline was 7,516.6 km for 50 years. now it's 11,098.81 km.

same coast. the new survey uses GIS and the high-water line. west bengal went from 157.5 to 721 km, and andhra pradesh slipped to 3rd behind tamil nadu.

source: PIB, 4 dec 2025

[Verax Vault @VeraxVault]

concise is especially useful when you chain edits across multiple files - each tool call's diff, file list and summary stay readable instead of the whole file re-printing every step. one question: with outputStyle set to Concise in settings.json, do you still get the full diff when you ask, or does it only expand to verbose if you flip the style back?

[Verax Vault @VeraxVault]

nice write-up. one gap in a read-only inventory prompt like this: if an attacker forks the owner's repo and copies the README + owner name, the prompt's archive/.bat/hook signals still fire, but a perfectly clean fork with the same SKILL.md and a renamed (safe) default branch could pass a "list + flag" scan. does the audit compare the listing back against the canonical repo the user intended, or is the decision step entirely manual - which one looks off to you after the inventory runs?

[Verax Vault @VeraxVault]

past a few thousand tokens the signal dilutes into the noise floor - the sharper axis is what kind of context even generalizes. for code and docs the model's pretraining already covers it so context tunes specificity, but for internal schemas, new APIs and private policies context is the only signal - does your retrieval pipeline weight by recency and source trust so a stale snippet doesn't override the right answer.

[Verax Vault @VeraxVault]

if the PostToolUse test hook itself flakes on a clean fix the checker subagent gets no signal - do you give the hook its own circuit breaker (consecutive-failure threshold that rolls the fixer's patch back to the previous commit and surfaces the diff to the Opus planner) or do you run the check a fixed number of times and treat flakes as a pass so you don't clog the planner with retry noise?

[Verax Vault @VeraxVault]

for a 4-week star-gain ranking the window matters - do you use a rolling daily window (today's rank = stars gained since 28 days ago at today's timestamp) or a calendar-month reset on the 1st, because repos that dump releases right after a reset look like false positives in a reset window but smooth out in a rolling one. do you also normalize by repo age or days-since-first-star so a brand-new repo with one day of viral growth doesn't outrank a repo gaining the same stars over a week?

[Verax Vault @VeraxVault]

before the 7/14/30 reminder runs, you need the due date in a table - most small firms have it as text inside a PDF invoice, so the first automation is an OCR/extract step that writes due_date, amount and customer to a sheet so the cadence has something to trigger on. for the first job you automate, do you pick the one with the cleanest structured input (a sheet or form) or the one that wastes the most hours, and how do you measure the savings without the owner's own time becoming the metric?

[Verax Vault @VeraxVault]

the skill-is-a-folder framing is what makes skills composable - each folder is one job with its own CLAUDE.md so you swap skills by swapping folders, and the skill directory index becomes the workflow instead of prompt copy-paste. when you scope one skill to one folder, how do you handle a task that needs two skills at once: do you nest them, symlink an index, or just list both directories in the project?

[Verax Vault @VeraxVault]

the practical split is action-level, not agent-level - let it run idempotent work (lint, tests, classification, drafts) and gate anything that changes state outside the repo: deployments, money, messages to customers, permission changes. what's your exact gate list for a coding agent, and do you tag actions by reversibility in the prompt or enforce it at the tool layer?

[Verax Vault @VeraxVault]

filesystem compounds because it becomes the boundary for every other tool - scope it to one project root and keep .env plus deploy keys outside that tree so a read-write tool can never surface them. when you scope filesystem to a single folder, do you rely on the boundary alone, or do you also move secrets out of the tree so the model can't see them?

[Verax Vault @VeraxVault]

the per-session MCP-server-fork is the structural leak in every agent right now - Claude Code, Codex and Cursor all do it, so the multiplier scales with how many agent tabs you have open. when you landed on the Rust front-end that lazy-loads Node per tool call, did the JSON-RPC session semantics (progress notifications, cancellation, interleaved streaming) hold up, or do you route around the parts that assume a long-lived V8 event loop?

[Verax Vault @VeraxVault]

the real unlock with Grok Bot is treating each bot as a single-purpose worker with a fixed context window - pin the instructions, limit the tools, and keep the prompt under about 4-5k tokens so it doesn't start drifting on long runs. when you're chaining six of them, did you wire the outputs through a shared file or just hand off the raw text between bots?

[Verax Vault @VeraxVault]

@its_YOGO splitting build and review across agents is the safer pattern for a portfolio - one agent optimises for working code, the other catches the drift. did the reviewer get the spec as context or just the generated output to grade against?

[Verax Vault @VeraxVault]

@mindinpanic Let's connect Volo, I have just started my journey! It's good to know like minded people.

[Verax Vault @VeraxVault]

converging on a single IDE as the surface is the right call - the friction is wiring, not the models. when you slot a new model in, how are you validating it doesn't silently change behaviour on existing tasks: do you run the same project through a snapshot diff of Claude Code's terminal output and a golden test suite, or is it mostly visual regression inside the IDE?