the-aios/aios/deep-research

Deep Research

Research for acting under uncertainty. The caller arrives with a provisional intent — "what should I write about?", "is X worth building?", "what's happening in this space?" — and this skill returns what to aim for, why, and how, grounded in a real survey of the field and the operator's own corpus.

The prime directive: intent is provisional

There is a chicken-and-egg at the heart of research: you need the research to know what you're actually looking for. A run that starts as "find my next post topic" may surface a product gap worth building, a decision that's become urgent, or a risk that reframes the week — and a run that starts as "should we build X?" may reveal the right move is to write about X instead.

So this skill treats the caller's intent as a starting lens, not a cage:

• Honor the lens — the caller's intent shapes where you look first and how you rank.
• Report what escapes it — when findings point at a different output mode than the caller asked for, say so explicitly. Never silently discard a discovery because it doesn't fit the original question.
• Re-examine the frame mid-run — after the sweep, ask: did the findings change what we should be aiming for? If yes, name the reframe before ranking.

Output modes

Every proposal names its recommended output mode — the form the finding wants to become:

One research run can return proposals in different modes. That's a feature — it's what makes the research honest.

The core discipline

Bad research dumps links. Good research triangulates, filters against what already exists, and privileges what's timely. Failure modes to design against:

• Dump — twenty sources, no synthesis, no ranking. The human still has to do the thinking.
• Duplication — proposing what the operator already wrote/built/decided. Map the corpus first.
• Evergreen bias — recommending timeless-but-cold moves while a perishable window is open. Timely > timeless when both are real.
• Lens-blindness — forcing every finding into the output mode the caller asked for. The mode is a finding, not a premise.

The loop

1. Frame — one line, held loosely

Write the provisional intent and what a good answer looks like: "{caller} wants {X}; a good answer is {shape}." This is the ranking filter — and step 4 explicitly revisits it.

2. Map the existing surface (BEFORE searching outward)

Read the operator's own corpus relevant to the domain — published work, drafts/pipeline, project notes, positioning/theses, prior decisions (Read/Grep over the relevant vault + repo paths). Inventory: what's been said/built/decided, what's half-done, and what the operator's distinctive angle or asset is. This turns "what's hot" into "what's hot that we are positioned to own" — and prevents proposing repeats.

3. Multi-angle sweep (never one query)

Run several WebSearch queries from genuinely different angles; WebFetch the 2–4 most promising sources for depth. Cover at least:

• The live conversation — what the domain is talking about right now, the phrases being used.
• Adjacent voices — what competitors / thinkers / publications the operator's audience reads are saying.
• Open questions — the contested, confusing, or unanswered points. White space lives here.
• Direction of travel — early signals of where the field is heading, not just current consensus.

Note publication dates on everything — recency is load-bearing for timeliness claims.

4. Re-examine the frame (the chicken-and-egg checkpoint)

Before ranking, ask explicitly: do the findings support the original intent, or point somewhere else? Three honest outcomes:

• Confirmed — the lens holds; proceed.
• Widened — the lens holds, AND findings surfaced moves in other modes; carry both forward.
• Reframed — the findings say the real aim is different; name the reframe as the headline and rank accordingly.

5. Triangulate + filter

Score each candidate on: timeliness (open window now? decaying by when?), fit (maps to an existing thesis/asset/draft that makes it fast, or genuine white space worth fresh effort?), and ownership (is the operator distinctively positioned, vs. anyone-could-do-this?). Drop duplicates of the corpus, cold-and-evergreen options when timely ones exist, and off-positioning moves.

6. Synthesize → ranked what/why/how proposals

Deliver 3–5 ranked proposals — a decision aid, not a link dump:

### {N}. {The move, stated as a move}
-**Mode:** write / build / decide / plan / act / watch
-**What:** {the concrete aim}
-**Why now:** {the timely reason — the open window, with source + date}
-**Why us:** {the operator's distinctive angle or asset, from the step-2 inventory}
-**How:** {the first concrete step + what it maps to (existing draft/repo/decision) or "fresh — needs {X}"}
-**Window:** {this-week / by-{date} / evergreen / watch-until-{trigger}}
-**Sources:** {2–3 links that ground it}

Lead with one line: "If you do one thing: {top pick} — because {reason}." If step 4 reframed the intent, lead with the reframe instead — that IS the finding.

End with what you could not confirm, and where you looked:

### Couldn't confirm
-{claim} — looked in {sources / queries, with dates}; {what would settle it}

Write "Nothing unconfirmed — every claim above is sourced" only when that is true. A verification-minded reader reads this block first: it is where a ranked list is most likely to be wrong, and a claim with no stated gap reads as a claim with no gap.

Rules

• Never a single search. Multi-angle sweep or it isn't deep research.
• Corpus first. You can't find white space without knowing what already exists.
• Intent is provisional. Honor the caller's lens; report what escapes it; reframe out loud when findings demand it.
• Every proposal carries what/why/how + mode. A topic without a mode and a first step isn't a proposal.
• Timely > timeless when both are real — and name the decay date.
• Cite dates. A "trend" from eight months ago isn't a window.
• Rank, don't dump. 3–5 proposals + a lead recommendation. A list of links means you're not done.
• Name the negative space. Every run ends with Couldn't confirm — the claims you could not verify and where you looked.
• Stop at the proposal. This skill chooses the aim; producing the deliverable (draft, code, memo) is the caller's or another skill's job.