Software Research

What this does

Turns one software-engineering question into a verified, multi-perspective deliverable. It picks a research mode, simulates that mode's expert lenses, maps where they contradict, synthesizes an HTML briefing plus a Markdown or ADR record, then adversarially verifies every claim against its primary source — version-aware — before delivering. Run the full pipeline; do not shortcut a phase.

Portability

Self-contained. Built-in tools only (Agent general-purpose, Write, web search/fetch inside agents) plus the files in this folder. Drop the folder into any .claude/skills/ directory and it works.

Phase 0: Scope & detect the mode

If $ARGUMENTS has the question, use it; else ask what to research.
Read data/modes.csv. Match the question against each row's signals (semicolon-delimited cues). Pick the best-matching mode_id.
- If two+ modes match with similar confidence, ask the user which mode (offer the matching modes + one-line descriptions). Otherwise proceed.
- If nothing matches, use deep-research (the fallback).
State the chosen mode + your one-line interpretation of the question. The user may override the mode.
Identify the reader role (developer / tech-lead / architect) from context; default tech-lead.
Derive a kebab-case topic-slug for filenames.

Gate 1 — confirm scope (one pause). Show a compact block and wait for the user to confirm or redirect before spawning any agents:

Mode:        <mode_id> — <one-line why this mode>
Question:    <your one-line interpretation>
Reader:      <developer | tech-lead | architect>
Lenses:      <the N lens names about to run in parallel>
Outputs:     <primary + secondary format, e.g. HTML briefing + MD>
Then:        adversarial verification against primary sources.
Proceed? (or correct the mode / scope)

This is the one place to catch a wrong mode or misread scope cheaply — before spending ~9-12 agents. Once confirmed, run Phases 1–3 autonomously (no more pauses until Gate 2).

Phase 1: Parallel expert lenses

Open references/modes/{mode_id}.md — it contains the exact lens prompts for this mode. Spawn that mode's lenses as general-purpose agents in a single message so they run concurrently. Give every agent the SAME question frame plus:

its lens prompt (from the mode file),
the source hierarchy + verification expectations from references/source-hierarchy.md (paste the tier list + all 7 rules — including Rule 7: load-bearing claims must resolve to Tier 1–3, never a blog/Q&A),
the instruction to return EXACTLY: (1) CORE POSITION in 2 sentences; (2) STRONGEST EVIDENCE, 3-5 bullets, each with a concrete data point + a primary-source URL and the version/date the claim applies to; (3) THE ONE THING only this lens would say. Under 400 words. Real fetched sources only — no invented studies, numbers, or URLs.

As each lens returns, append its brief to a working scratch file software-research-reports/{topic-slug}.work.md (create it now) under a ## Lens: <name> heading. Writing briefs as they land means a crash mid-run loses nothing — the run is resumable from what's already on disk. Set the run-state block at the top of that file: stepsCompleted: [0, 1].

When all return, post a 2-3 line note in chat: convergence + sharpest disagreement. Keep raw briefs out of chat.

❌ Common failures here:

Citing a blog / SO answer / model memory as proof instead of climbing to the primary source behind it (Rule 3 + Rule 7). The blog locates the fact; the primary is the fact.
Presenting an unversioned claim as current — "X does Y" with no version is unverified (Rule 1).
Letting one loud lens set the narrative before the others return; wait for all N.

Phase 2: Map the contradictions (inline, no agents)

From the briefs only, determine:

Direct conflicts — name the specific clashing claims.
Strongest vs weakest evidence — rank by source tier (Tier 1 primary/spec > Tier 2 security/aggregator > Tier 3 registry > Tier 4 benchmark > Tier 5 survey).
The resolving question — the single empirical test that settles the biggest conflict.
Universal agreement — what every lens confirms (load-bearing finding).
The blind spot — what NO lens addressed (missing 6th lens → frontier question).

Append this map to the scratch file under ## Contradiction map and bump the run-state to stepsCompleted: [0, 1, 2]. The map is the raw material for the synthesis; not a separate deliverable.

Phase 3: Synthesize the output(s)

Read the mode's primary_output and secondary_output from data/modes.csv.

Before writing any output, load references/report-structure.md. It defines the concreteness contract (specific > generic: exact versions, real commands/config/ code, real numbers, tied to THE READER's stack — and a ban list of filler phrases) and the 12-section deep spine. Detail is not optional padding: a report earns its length by being practical and about the reader's situation, not a survey of the topic.

For html: clone assets/briefing-template.html; do not rebuild the CSS. Fill every {{TOKEN}}. It has two layers: (a) the fast-scan layer up top — verdict scoreboard, key findings, and (for option-comparison modes) the Side-by-Side table — so the answer lands in five seconds; then (b) the 12-section deep body (sections 06–16) that carries the full spine. Fill the deep sections per the concreteness contract; render comparisons/numbers/flows with widgets, not prose.
For md: the full 12-section spine from references/report-structure.md — introduction/methodology, landscape, implementation how-to, stack, integration, performance, security, recommendation, roadmap, risks, methodology, and a closing References appendix. This is the detailed report; make each section concrete. In the References appendix every source is a clickable [title — version/date](url) Markdown link (never a bare URL in a table cell — those don't reliably render as links).
For adr: clone assets/adr-template.md; fill the MADR sections (Context, Decision Drivers, Considered Options, Decision Outcome, Consequences, Pros/Cons per option, More Information) and the closing References section. Document rejected options + why. The ADR stays a lean decision record — no 12-section spine — but the concreteness contract still applies (name versions, cite specifics).

Write for a busy engineer, not a journal. Short sentences, plain words, lead with the answer. Define any unavoidable term inline. If a reader needs a dictionary to parse a finding, rewrite it.

Add a visualization when it earns its place. assets/widgets/ holds ~45 drop-in, zero-dependency visual blocks (inline SVG / CSS / a little vanilla JS). Read assets/widgets/README.md — it has the full catalog grouped by job (compare & decide, numbers & evidence, explain how it works, sequence over time, org & management, inline accents) plus a numbered "which visual to use" decision order. Pick by what the content is, not by novelty; copy the file, replace its FILL: markers. A few high-use anchors: comparison-matrix (options × criteria), weighted-decision-matrix (ADR scoring), metric-bars (benchmark/cost/size), evidence-confidence (per-finding trust), callout (gotchas/warnings), verdict-card (the recommendation), and slider-metric (the one interactive "drag it" widget — at most one per report).

Default to the zero-dependency widgets: they render instantly, work offline, print, and can't be blanked by a CDN outage or a runtime handshake failing. For any tree/graph/diagram the widgets use hand-placed coordinates — don't auto-layout. Reach for a library only for what bespoke SVG is bad at: a standard flow/sequence/ER diagram easier written as text → Mermaid (pinned, securityLevel:'strict'); real quantitative data at scale → Chart.js; code snippets that need syntax coloring → code-compare (it loads highlight.js pinned+SRI — include that block once per report, set class="language-…" per snippet). Then pin the version + prefer SRI/inline over a bare CDN <script>. If you let the model generate arbitrary widget JS rather than filling a template, sandbox it (iframe sandbox="allow-scripts" without allow-same-origin, CSP blocking connect-src) — generated markup is untrusted code in the report's origin; SRI does not help there. Full rules in assets/widgets/README.md.

Write outputs to software-research-reports/ (create if needed):

HTML → {topic-slug}-briefing.html
MD → {topic-slug}.md; ADR → ADR-{topic-slug}.md

❌ Common failures here:

Burying the recommendation under prose — the scoreboard and Bottom Line exist so the reader gets the answer before the detail.
Academic / oblique phrasing that reads as clever but doesn't inform.
For an X-vs-Y question, skipping the comparison table (the reader wants the side-by-side).

Phase 4: Adversarial verification (mandatory)

Gate 2 — confirm findings before verifying (one pause). With the draft synthesized, post a tight list in chat: the recommendation, the 3–5 key findings, and which claims are load-bearing (the recommendation rests on them). Ask the user to confirm the finding set / flag anything to scrutinize harder. This is the last cheap moment to redirect before spending verifier agents. Then run 4a–4c.

verify_depth comes from data/modes.csv (full = every citation; load-bearing = the claims the recommendation rests on — still mandatory).

4a. Self-review (inline). Score each finding 1-10 for reliability (by source tier, not confidence) and justify. Identify the weakest link + what would verify it. Bias check: which lens dominated, what got underweighted. Name the missing 6th lens. Assign an honest overall grade.

4b. Verify citations (parallel agents). Spawn general-purpose agents in one message, one per citation cluster (~4-6). Each prompt: independently verify the claim against its PRIMARY source, applying all 7 rules in references/source-hierarchy.md (version-bind; check current-version validity; climb to primary; date-stamp; security via OSV/GHSA + version range; benchmarks need reproducible methodology; load-bearing claims must resolve to Tier 1–3). Return VERDICT = CONFIRMED / PARTIALLY CONFIRMED (list corrections) / UNVERIFIED / FALSE / VERSION-STALE / UNVERIFIED-LOWTIER (a load-bearing claim backed only by Tier 5–6 — no primary found), the corrected one-line citation with version+date, and 2-4 specifics with the primary URL. Under 280 words.

4c. Apply corrections. Fix wrong figures/titles/dates/versions. Downgrade confidence where evidence is thin; demote contested/preprint/version-stale AND UNVERIFIED-LOWTIER claims into the contested sidebar (a load-bearing claim with no primary backing must not stand as a finding). Fill the verification banner (N checked · X corrected · Y demoted · Z version-stale) and per-citation status tags. Populate the claim-safety guide (assert / caveat / avoid), the version-currency note, and confirm every source appears in the References section. Bump the run-state to stepsCompleted: [0, 1, 2, 3, 4].

❌ Common failures here:

Skipping verification because the findings "look plausible" — plausible-but-wrong is exactly what an LLM panel produces; the verifier is the guard.
Leaving a load-bearing claim backed only by a blog/SO answer in the main findings instead of demoting it (Rule 7 → UNVERIFIED-LOWTIER).
Verifier "confirming" from memory instead of fetching the primary source.

Output

Deliverables: the post-verification HTML + MD/ADR in software-research-reports/.
Open the HTML with the platform opener: macOS open <path>, Linux xdg-open <path>, Windows start "" <path>. If unclear, print the path.
Chat summary: file paths; verification tally; the one universal finding; the recommendation + its load-bearing claim; the frontier question; the claim-safety summary (safe to assert vs avoid). Keep it tight.

Notes & guardrails

Web search is required. The lenses and verifiers depend on fetching live primary sources. If web search/fetch is unavailable inside the agents, abort and tell the user — never answer version- or security-sensitive software questions from training data alone; a model's memory is Tier 6 and goes stale.
Real research only. Every lens and citation traces to a real, fetched primary source with a version/date. If a figure can't be verified, demote or cut it; never paper over it.
Prefer official docs over Q&A. A load-bearing claim must resolve to Tier 1–3 (official docs/specs, security DBs, registries). Blogs, Stack Overflow, Medium, and model memory are signposts to find the primary — never proof on their own.
Specific and practical, never generic. Name exact versions; show real commands/ config/code; use real numbers; tie every recommendation to the reader's stack, team, and scale. Cut any sentence that describes a category instead of stating a fact ("offers robust support for…", "is widely used…", "can help improve performance…"). If a sentence wouldn't change what the reader does, delete it. Full contract + good/bad example in references/report-structure.md.
Scratch file. {topic-slug}.work.md holds lens briefs + the contradiction map for resumability; it is a working artifact, not a deliverable. Leave it in place (a resumed or re-run pass can reuse it); the deliverables are the HTML/MD/ADR.
Version-aware. "X does Y" is only valid as "…in version N". Flag deprecated or version-stale claims; never present stale info as current.
The panel is author-built. Disclose it. Lens agreement is a strong hypothesis, not field consensus.
Reliability = source-tier evidence quality, not confidence.
Cost. ~9-12 agents per run (lenses + verifiers). Expected. Don't fan wider than the mode's lenses / one verifier per citation cluster.
Design. Keep the HTML template CSS verbatim (clean white, Montserrat / Roboto Mono, blue accent).