research-grounded

name	research-grounded
description	Audit a plan's design decisions — library, framework, pattern, and architecture choices in IMPL.md / ROADMAP.md — for missing `# research(YYYY-MM):` provenance, then web-search to ground them. Use when you want architectural choices verified against current best practice before they harden into code, rather than resting on training-cutoff recall. Catches committed decisions stated as fact but never checked — the class of miss that becomes revertable work.
disable-model-invocation	false
allowed-tools	Bash Glob Grep Read Edit WebSearch WebFetch Agent

Research-grounded

Catch design decisions that were made without checking current best practice — before they become code. A planning doc that says "we'll use X" with no provenance is a bet on training-cutoff recall; the convention # research(YYYY-MM): <tradeoff> <source> records that the bet was checked against reality. This skill finds the bets that weren't.

Repo context

/techne:research-grounded audits docs in the CWD repo by default, or a path you name (which may live in another repo). Resolve the repo root from the argument:

git -C "$(dirname "<target-doc>")" rev-parse --show-toplevel   # no path arg → CWD repo root

The target's .claude/skill-context.md isn't required, but its ## repo section helps you tell a genuine technology choice from incidental prose.

What needs provenance

A committed design decision — a choice with real alternatives, where picking wrong is expensive to undo:

• Library / framework / tool — "MapLibre + PMTiles", "Playwright over Cypress", "switched from JWT to server sessions".
• Pattern / technique — "event delegation over per-node listeners", "stale-while-revalidate", "container queries for card internals".
• Architecture / protocol — "A2A Message.metadata instead of text tags", "OPFS for tile storage".
• Version / API-surface bets stated as fact — "ElevenLabs v3 supports SSML break tags" (the miss that cost 5 revertable PRs: a capability asserted, never verified against the target's docs).

Each of these, when committed in IMPL/ROADMAP, should carry an adjacent # research(YYYY-MM): <one-line tradeoff> <source>. Negative-space decisions ("we did not do X because…") need provenance too.

Ignore

• Descriptive implementation prose. "reads the metrics object instead of re-reading constants", "returns a structured fallback instead of a retry loop" — uses decision words but picks no technology or pattern. Not research-worthy.
• Hypotheticals & aspirations. "what if we used sessions instead of JWT", "we plan to", "coming soon" — not yet a commitment (mirrors how /techne:docsync waves through roadmap language).
• Already grounded. A decision with a research(...) tag on the same bullet or within a few lines — even a terse one.
• Bugfixes, renames, mechanical changes — no alternative space to research.
• Settled house conventions — the fleet's own recorded standards (SHA-pinning, squash-merge). Don't re-flag every mention.

Workflow

1. Scope. Default: IMPL.md + ROADMAP.md in the target repo. Accept a named file or a directory (e.g. planning/). For multiple files, fan out — one Explore subagent per file (see Fan-out).

2. Find candidate decisions — two grep passes, then read the design sections. The greps seed candidates; judgment decides which are research-worthy per the lists above.

   Pass A — decision verbs (catches "Library … over …" and "Pattern over pattern" phrasing):

   grep -nEi 'chose|switched|adopted|replaced|opted for|we use|picked|migrated|over [A-Za-z.-]+ \(|instead of' ROADMAP.md IMPL.md
   

   Pass B — capability & best-practice-as-fact (catches the costliest class — "Version / API-surface bets stated as fact" — which carries no decision verb, so Pass A is blind to it):

   grep -nEi 'supports?|exposes?|provides?|enables?|natively|built-in|out of the box|standard|canonical|idiomatic|recommended|modern|best practice' ROADMAP.md IMPL.md
   

   Pass B surfaces "RealtimeTTS exposes word-timing for Kokoro", "Standard FastAPI 2026 setup", "the 2026 modern … standard is X" — claims a verb-grep never sees.

   Neither grep is sufficient — also read the doc's design / architecture sections directly. Both passes are noisy and leaky: a bet phrased as a bare proper-noun list ("Postgres + SQLModel + asyncpg", "MapLibre + PMTiles", "on Message.metadata") trips no keyword at all. The real signal is proper-noun library / protocol / API names; .claude/skill-context.md's ## repo section tells you which tokens are this repo's technology choices. Classify per-hit, not per-match — most Pass A instead of hits and most Pass B supports/standard hits are descriptive prose.

   research(2026-05): off-the-shelf prose linters (proselint / write-good / Vale) flag style weasel-words ("very", "quite"), not capability-claims-missing-provenance — so this seed is bespoke by necessity, not a reinvented wheel.

3. Check provenance. For each genuine decision, look for a research(YYYY-MM): (or # research:) tag on the same bullet or within a few lines. Present → grounded, skip. Absent → gap.

4. Report gaps. One entry per un-grounded decision:

   ROADMAP.md:88
     decision: "embed offline maps with MapLibre + PMTiles"
     gap:      no research provenance — grounded in 2026 offline-map options, or recall?
     action:   web-search current PMTiles/MapLibre tradeoffs → add `# research(2026-05): <tradeoff> <source>`
   

5. Ground or flag (on confirmation).

   • Ground it (the point of the skill — close the loop, don't just flag): web-search the decision's current best practice, then add # research(YYYY-MM): <tradeoff> <source> capturing what you actually found.
   • Flag only: if grounding needs the user's domain context a quick search can't supply, leave a one-line note to ground it before it hardens into code.
   • Never invent a citation. A research() tag with a fabricated or unread source is worse than no tag.

Fan-out pattern

For a directory (e.g. planning/), spawn one Explore subagent per file; each returns a gap report for its file. Consolidate, present grouped by file, then ask ground all / ground selected / skip?.

Don't touch

• Docs the user didn't name or imply; git history / changelogs (historical, not live bets).
• research() tags that already exist — don't reword a grounded decision unprompted.
• Decisions already shipped as code with no alternative left to research — audit the plan, not the past.

Why this skill is quiet

Output is the gap report and, on confirmation, the web-searched research() tags. No narration of the scan — just decision → gap → action. Sibling of /techne:docsync (claims vs code); this is decisions vs evidence.