Jonathon's AI Wiki

Nine AI Techniques Worth Trying (Late 2026)

10 min read

Source: raw/9_AI_Techniques_You_Probably_Haven_t_Tried.md — The AI Daily Brief (Nathaniel Whittemore), ~30 min (https://www.youtube.com/watch?v=o5PzgHXXjjQ)

A deliberate aggregation rather than a first-person report — the host is explicit that this is “not my experience with specific techniques… but instead an aggregation of what other people are sharing about how they’re learning.” Useful as a scan of what changed while you were heads-down, and as a pointer list. Each item below records who is actually making the claim.

Key Takeaways

  • Voice is the sleeper. Codex’s live voice mode is drawing the strongest reactions in the roundup, and the argued reason is not any use case — it is ambient interaction. “It probably sounds like a negligible difference to speak to an ambient assistant versus click-to-speak, but it is a world of distance in practice” (Ali K. Miller).
  • You can now teach by being watched. ChatGPT’s computer history is the ambient version (it keeps a timeline of how you work, which you turn into repeatable processes); Grok Bot’s watch-this-task button is the deliberative version (you say when to watch and for how long). Tasks previously too tedious to explain may now be demonstrable.
  • The anti-AI-ism fix is a skill, not a re-prompt. If you find yourself typing “rewrite this without all the AI-isms,” capture the tells once and install them. Ruben Hed’s list of giveaways is offered as a paste-in seed — and it names a failure the wiki has not recorded before: the model clapping for itself (“And that matters. That’s the part everyone misses”).
  • Skills are a discipline, not a switch. “Knowing where to insert skills, which skills to give it, where to back off and just let it work natively — all of those become really important agent-management capabilities.”
  • Multiplayer AI is the named trend for the fall. Agents that live where teams meet rather than on one person’s machine. Claude Tag is the cited best example; the argument is that work is collaborative, with handoffs and shared context, and single-player agent setups leave that on the table.
  • Local AI crossed a usability line. Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index — “state-of-the-art just a few months ago” — while running on common hardware.
  • The host’s honest meta-point: most of these “aren’t necessarily some massive change to your workflow, but just some simple little ideas to get a bit more out of the tools that you now use every day.”

The nine

1. Voice mode

Codex’s live voice mode, mostly via Ali K. Miller, who describes it as “an ambient workforce, an ambient voice-controlled operating system” and reports finishing urgent work at 2am because of it. Her examples are conversational rather than task-shaped: “Take a look at this paragraph, what do you think?” / “I’m going to edit this while you find that doc.” / “Just sent the 15 client emails — check against the list while I start on the script.” She uses it on walks and as a weekend background process. Dan Shipper (Every) reports the same: “almost every single person at Every is freaking out about how good voice mode in ChatGPT for Work is.”

The host’s caution is the useful part: this is not something you evaluate from a description. Commit to a period of working this way even though it feels weird, then decide.

2. Teach the agent your workflow by letting it watch

Two shapes, as above — ChatGPT computer history (ambient timeline) and Grok Bot’s explicit watch mode (deliberative). The reframe offered: if you previously ruled a task out because explaining it was harder than doing it, that constraint may have moved.

3. Make a skill that removes your AI tells

Aimed at the volume writing everyone already outsources — emails, memos, summaries, website copy. The named tells: tiny compressed sentences of a word or two separated by periods for drama; “it’s not this, it’s that”; the performative honest caveat; and clapping for itself. The mechanic offered is deliberately crude — print the list of tells as a PDF, drop it into Claude, ask it to make a skill.

The roundup also notes people converting formal style standards into skills, naming ASD-STE100 (the Simplified Technical English standard used for aircraft manuals). This wiki has an unresolved contradiction on ASD-STE100 — recorded as tried-and-rejected in one source and working in another — where the difference appeared to be the injection point. A skill is a third injection point, distinct from both. See Instruction Injection Layers.

4. /design in Claude Code

The host’s framing adds something to the existing coverage — it is an improvement at two different scales:

  • Macro: when you do not know how you want a thing to work, having it lay out several templates you can view at once and react to at a high level.
  • Micro: editing a specific part of a design instead of re-prompting the whole thing and losing everything you liked.

5. Skills as an ongoing practice

Points at Matt Pocock’s catalog at aihero.dev/skills, and the organising principle is the transferable bit: skills grouped by when you would use them — getting started, main flow, shaping, upkeep — rather than by what they do.

The exemplar named is grill-with-docs, a skill that interviews you about a plan or design until you and the agent share one understanding, and writes the vocabulary and the hard decisions into your repo while it does. The host’s emphasis: “importantly, it is not just an interview, but an interview that comes with a leave-behind.” See Context-Extraction Skills — and note this wiki has carried Pocock’s repo since May 2026, so the news here is the pattern’s reach, not its invention.

6. Multiplayer AI / team agents

The argument: when OpenClaw arrived you became the manager of a team of agents, but they were your agents. Real work has handoffs and shared spaces. The prediction is that the important tools this fall will be the ones that live where teams come together.

Claude Tag is the cited example, and the distinction drawn against the older Slack integration is precise: Claude joins as a team member with access to the whole channel’s context, and that channel’s Claude can have its own permissions, tool access, and context — different from Claude in another channel — while being a shared resource rather than something on one person’s laptop.

Second example: 10X’s “citizen SDLC,” a six-stage lifecycle for taking what non-technical employees build with AI from a personal prototype to something the whole company runs on. Enterprise vibe-coding, promoted to production.

7. Grok Bot

Covered in depth in Grok Bot. The two points unique to this source: bots can already interact with one another, which makes a chief-of-staff topology possible, and the host expects shared-across-a-team bots next. The other: Nous Research shipped “bot mode” for Hermes desktop, described here as very similar to Grok Bot but on an open, more controllable, more customisable stack — the option for anyone who does not want to work with Grok.

Also named: Lenny Rachitsky’s idea of pointing a bot at an MCP holding transcripts of 500+ podcast episodes to make a product-strategy, growth, and career adviser. The generalisation offered — wherever you interact with a lot of information in a regular, predictable way, this shape fits.

8. Local AI

Qwen 3.8 27B, at 52 on the Artificial Analysis Intelligence Index, running on common hardware. See Qwen 3.8 27B for the four-source picture including measured throughput on a Mac.

9. Two-word prompts

Ali K. Miller’s list of 18, of which four are described. They read like things you would say to a colleague, which the host connects back to voice mode:

  • “now what” — after wrapping a project or a big push, when you still have energy
  • “please fix” — usually with a screenshot flagging the issue
  • “simulate it” — run scenarios, plan for edge cases, and turn the result into a planning interface
  • “remember this” — force a mistake or a critical context error into memory. Her note: the model often does this automatically, but not always, and being specific helps.

Also in this episode (headlines segment)

Two items with consequences beyond the roundup:

  • OpenAI’s “private safety processing.” For eligible API customers with zero-data-retention agreements, safety scanning is extended across an entire session — including customer-controlled context storage and multiple agentic steps — while staying fully automated and encrypted, so no human sees the data. Flagged issues reach an OpenAI employee as a summary with a category and severity rating, stripped of customer data. The stated motivation is that long-horizon agents broke per-interaction monitoring. The comparison drawn is directly relevant to model routing here — see Anthropic’s Position.
  • Replit “free mode” routes all queries through GPT-5.6 Luna for $20/month users, which Replit says allows ~30× more creation on the same subscription, with a manual escalation to a stronger model. The host’s read: OpenAI is now competing on the efficiency frontier as deliberately as on the capability frontier, and expects more of it “as we see more maturation among users in terms of their understanding of which power levels are required for which different types of tasks.”

Try It

  1. Pick the one item you have not tried and block an afternoon. The host’s whole thesis is that these get evaluated by use, not by description — voice mode especially.
  2. Write down your own three AI tells before installing anyone else’s list. The generic ones are real, but your reviewers notice yours.
  3. Add “remember this” to your vocabulary the next time you correct a persistent context error, rather than assuming memory captured it.
  4. If you run a team, price out the multiplayer question now. The distinction that matters is per-channel permissions and context, not just “the agent is in Slack.”

Open Questions

  • This is aggregation, not testing. Every claim here is someone else’s report, and the host says so up front. Nothing in the roundup is measured.
  • ChatGPT computer history mechanics. Described as an ambient timeline you can turn into repeatable processes; no detail on retention, scope, or controls. Given the deletion-semantics finding in What Personal Agents Keep, this warrants its own look.
  • Hermes “bot mode” is a one-line mention. Similarity to Grok Bot is asserted, not demonstrated. Needs first-party confirmation from Nous.
  • 10X’s citizen SDLC is named with a six-stage structure but the stages are not listed. Worth a separate ingest.
  • Ali Miller’s other 14 two-word prompts are not enumerated in the episode.

Graph View

Backlinks