agentleFS
rubiconrubicon/ai-skills/skills/work-evidence-research/SKILL.md

work-evidence-research

Skill1 starsChanged 3 months agoMITView on GitHub

Forensic, source-grounded research workflow for reconstructing work history and client proof from connected files (Dropbox, Google Drive, and similar). Runs a three-pass workflow — discovery, targeted verification, final assembly — and classifies findings by client, firm/era, project type, evidence type, and outcome type, while keeping employer-side and agency/client-side work separate, quarantining unsupported claims, and never inventing details. Use when the user wants to build a client evidence database, find all projects of a given type, verify which projects have real outcome metrics, separate employer-side from agency/client work, assemble case-study candidates, or produce a fully-quoted CSV source appendix. Trigger phrases: "build a client evidence database," "find all website projects," "reconstruct my work history," "which projects have real outcome metrics," "separate employer-side from agency work," "source appendix CSV," "case-study candidates," "final assembly using prior findings."

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 7 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Source-aware: Every factual claim must be traceable to an actual found file.

    line 39
  2. Distinctive:

    If source linkage is uncertain, mark it explicitly — do not smooth it over.

    line 40
  3. Very specific:

    No invented details: Never fabricate clients, dates, titles, metrics, or relationships.

    line 41
  4. Distinctive:

    Preserve ambiguity: Keep unresolved items visible (probable vs confirmed, alias conflicts, quarantined claims).

    line 44
  5. Distinctive:

    Separate context: Keep employer-side work (done as an employee for the company) and agency/client-side work (done for an external client) …

    line 46
  6. Distinctive:

    Scope discipline: Search only what the user asked about.

    line 54
  7. Distinctive:

    Use this skill when the user wants to:

    line 60
  8. Very specific:

    Do not restart broad discovery when prior findings already exist.

    line 104
  9. Distinctive:

    If no file source is connected, ask which to search rather than inventing.

    line 135

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Work Evidence Research

A forensic research workflow for reconstructing work history and client proof from a person's connected files (Dropbox, Google Drive, and similar). It treats the file corpus as evidence: it discovers candidate work, classifies it by strength, keeps employer-side and client-side work apart, quarantines what it cannot prove, and assembles reusable deliverables (evidence databases, case-study shortlists, resume bullets, source appendices). It is a research method, not a summarizer — every claim must trace to a found file.

Evidence standards (read first)

These are non-negotiable. They override any instinct to produce clean, polished, or complete-looking output.

  • Accuracy over polish. A smaller, defensible result beats a fuller one that guesses.
  • Source-aware. Every factual claim must be traceable to an actual found file. If source linkage is uncertain, mark it explicitly — do not smooth it over.
  • No invented details. Never fabricate clients, dates, titles, metrics, or relationships. Never invent or reconstruct URLs, file paths, dates, or metadata. Where a fact is missing, use a clearly-marked [FILL IN] placeholder.
  • Preserve ambiguity. Keep unresolved items visible (probable vs confirmed, alias conflicts, quarantined claims). Do not force premature resolution.
  • Separate context. Keep employer-side work (done as an employee for the company) and agency/client-side work (done for an external client) separate, even when the same company appears in both — unless direct evidence justifies merging.
  • Quarantine, don't upgrade. Claims that lack a supporting artifact go in a visible quarantine; never silently upgrade or omit them.
  • Later corrections win. A user correction overrides earlier assumptions and earlier drafts. Record the override; preserve the conflict rather than collapsing it.
  • Scope discipline. Search only what the user asked about. Do not wander into unrelated files, and do not restart broad discovery when prior findings already exist.

When to use

Use this skill when the user wants to:

  • Build a client/work evidence database for a firm, era, or set of clients.
  • Find all projects of a given type ("all website projects," "all analytics work") across one or more eras.
  • Verify which projects have real outcome metrics, or close gaps on probable finds.
  • Separate employer-side from agency/client-side work for a company.
  • Assemble case-study candidates, a resume bullet bank, or portfolio-safe blurbs.
  • Produce a source appendix — a fully-quoted CSV ledger of the files used.

Entry dimensions

A request can enter from any of five dimensions; identify which one(s) apply:

  1. Client / brand — a named client or a list (with aliases).
  2. Firm / era — a company worked for and a time range.
  3. Project type — treated as a concept cluster, never a single keyword ("website" → redesign, launch, CMS migration, sitemap, IA, UX/UI…).
  4. Evidence type — the kind of source. It spans document genre (proposal, SOW, deck, campaign report…), artifact role (primary proof / corroborating / mention), and file format (.pdf, .docx, .pptx, .csv…). Format aids retrieval only; genre, role, and tier drive classification.
  5. Outcome type — the result language sought (revenue, conversion, traffic, leads…), kept distinct from scope/operational, client-profile, and awards metrics (see the full metric taxonomy).

See references/taxonomies.md for the clusters, genres, formats, tiers, and metric rules.

The three-pass workflow

  1. Discovery — broad inventory across the connected files for the chosen dimension(s); classify each find by evidence tier; flag gaps.
  2. Targeted verification — for probable / needs-verification items, hunt the specific missing artifact; upgrade a tier only on a direct artifact, downgrade when a claim fails to verify.
  3. Final assembly — produce the polished deliverable from confirmed findings plus prior context.

Mode selection comes first: detect whether the user wants fresh discovery, verification of prior findings, final assembly from established research, or a source appendix — and route accordingly. A request may combine passes (discover then verify; find-by-type then filter by outcome) — run every pass it implies, in order. Do not restart broad discovery when prior findings already exist. See references/workflow-modes.md.

Classification & metrics

Classify every find into an evidence tier (Confirmed / Probable / Needs verification / Role/employer / Duplicate/alias) and separate true outcome metrics from scope, client-profile, awards, and quarantined claims. Tier is set by artifact role and content, never by file extension or genre label. Full taxonomy in references/taxonomies.md.

Output modes

Deliverables (skeletons in references/output-formats.md): client evidence database, confirmed/probable watchlist, quarantined-claims table, alias-resolution table, resume bullet bank, case-study candidate ranking, and a fully-quoted CSV source appendix. For large output, chunk it (one labelled part at a time, wait for "continue"). When emitting a CSV appendix, follow the hard no-invention rules in that file.

How to handle a request

  1. Identify intent — relationship proof, project discovery, case-study selection, metric verification, source-appendix generation, or final packaging.
  2. Identify the entry dimension(s) and pass, and select the mode (fresh vs verify-prior vs assemble-from-prior).
  3. Open the relevant reference file(s) and work from them.
  4. Gather inputs — target firm/era, clients, aliases, project types, corrections/constraints, prior context, preferred output mode. Ask concise questions only when missing information blocks progress (e.g. no file source is connected, or the target firm/era can't be identified); otherwise proceed on the most defensible interpretation and preserve the ambiguity rather than stalling. If no file source is connected, ask which to search rather than inventing.
  5. Run the pass(es) and produce the deliverable in the requested output mode.
  6. Preserve the user's own facts and the messy edges — quarantine, aliases, conflicts — rather than smoothing them away.

Examples

  • "Build a client evidence database for [FIRM], 2015–2019." → discovery across the firm/era, tiered classification, employer-vs-client split, database output.
  • "Find all website projects from [FIRM]." → expand the website concept cluster (and the .html / .pptx / .docx format lane), classify confirmed vs probable.
  • "Which [CLIENT] projects have real outcome metrics?" → outcome-language search, separate outcomes from scope, shortlist case-study candidates.
  • "Separate employer-side from agency-side work for [COMPANY]." → two distinct context rows, no merged metrics, uncertainty preserved.
  • "Make a fully-quoted CSV source appendix for everything used." → source-appendix mode, every field quoted, NO_URL / UNAVAILABLE for missing data.
  • "Final assembly from what we already found — don't re-discover." → assembly mode only, no broad rediscovery.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.