Distill Industry

Distill an Industry 2.0

Produce two linked but separate products:

1. a dated, auditable industry decision pack for people; and
2. a capability map plus tested atomic Agent Skills for recurring work.

Keep changing facts in the evidence pack. Put only durable procedures, decision rules, schemas, and quality gates in Skills.

Select the mode

• scan: orient quickly, expose unknowns, and decide whether deeper research is justified.
• full: build the complete industry pack, opportunity map, entry playbook, capability map, and Skill candidates.
• skill-pack: create Skills from an existing evidence pack and capability map.
• refresh: update expired or changed claims and rerun affected decisions and tests.
• compare: apply identical boundaries, evidence rules, and scoring to multiple industries.
• audit: inspect an existing industry pack or Skill Pack without rebuilding it.

Default to full only when the request clearly asks for a complete system. Otherwise choose the smallest mode that satisfies the decision.

Run the 11-stage pipeline

Read [references/research-workflow.md](references/research-workflow.md) before full, compare, refresh, or skill-pack work. It defines stages, artifacts, gates, continuation rules, and readiness limits.

1. Resume or scope — read PIPELINE_STATE.md when present; otherwise freeze industry, exclusions, geography, customer, user role, decision, date, depth, sources, and constraints in SCOPE.md.
2. Plan questions — create QUESTION_MATRIX.csv from [assets/question-matrix.template.csv](assets/question-matrix.template.csv). Give every material conclusion an evidence route.
3. Extract independently — research all required lenses. Use isolated workers only when the host supports them; otherwise execute the same task packets sequentially with clean role boundaries.
4. Normalize evidence — split compound claims and write one source-to-claim relationship per ledger row using [assets/evidence-ledger.template.csv](assets/evidence-ledger.template.csv).
5. Pass V1 Evidence Validation — authenticate scope, freshness, independence, locators, counterevidence, conflicts, and claim-to-source relations.
6. Synthesize the system — create the overview, value chain, market model, customer/competition analysis, economics, risks, and unknowns.
7. Pass V2 Decision Validation — reproduce calculations, stress assumptions, test alternatives and falsifiers, and limit conclusions to the available evidence and user fit.
8. Map capabilities — build the industry capability map before selecting Skills. Use [references/capability-map-standard.md](references/capability-map-standard.md) and [assets/capability-map.template.csv](assets/capability-map.template.csv).
9. Select and construct Skills — record every candidate in [assets/skill-candidate-register.template.csv](assets/skill-candidate-register.template.csv) as accepted, deferred, or rejected; build accepted Skills under [references/skill-pack-standard.md](references/skill-pack-standard.md) using [assets/atomic-skill.template.md](assets/atomic-skill.template.md).
10. Link and pass V3 Execution Validation — map depends-on, composes-with, contrasts-with, and supersedes; run positive, negative, edge, sibling-confusion, missing-input, execution, degradation, safety, and blind tests.
11. Deliver and grade — produce the executive brief, index, validation report, limitations, refresh triggers, and next experiment. Install or publish only when the user requested that external action.

Update PIPELINE_STATE.md after every completed stage. Never restart a valid completed stage without a recorded reason.

Apply the three validation gates

Read [references/triple-validation.md](references/triple-validation.md) before assigning any readiness level.

• V1 Evidence: fact and evidence integrity.
• V2 Decision: calculations, commercial reasoning, alternatives, and action thresholds.
• V3 Execution: triggering, workflow adherence, output, safety, degradation, reuse, and independence.

For each gate report passed, partial, failed, or not-tested, with evidence and failure reasons. Enforce these limits:

• V1 failed: do not issue a commercial recommendation.
• V2 failed: label conclusions as research hypotheses; do not issue an entry decision.
• V3 failed or not tested: do not claim generated Skills are stable or pilot-ready.
• Any critical gate below passed: cap overall readiness at draft.

Control evidence

Read [references/evidence-standard.md](references/evidence-standard.md) before collecting or rating sources and [references/analysis-frameworks.md](references/analysis-frameworks.md) before market, value-chain, economics, opportunity, or risk analysis.

Classify claims as verified, supported, tentative, disputed, or unknown. A verified material claim normally needs two independent evidence origins and a primary or authoritative source when reasonably available. Never count syndication, a vendor report and its derivative article, or repeated pages from one dataset as independent evidence.

If live research is unavailable, produce a source-constrained result, list missing evidence, and preserve unknowns. Never invent citations, current figures, tools, or verification work.

Generate a complete but non-inflated Skill Pack

Do not target a fixed Skill count. Demonstrate coverage before selection across:

• industry intelligence and monitoring;
• customer discovery and demand validation;
• prospecting, qualification, and prioritization;
• solution, pricing, economics, and decision support;
• transaction, negotiation, procurement, and handoff;
• delivery, operations, quality, compliance, and safety;
• diagnosis, optimization, renewal, and learning;
• orchestration only when the end-to-end request recurs.

For each material lifecycle gap, either create a candidate or record why no reusable Agent task exists. Do not convert report chapters, generic advice, volatile facts, or expert personas into Skills.

An accepted candidate must be repeatable, actionable, distinct, grounded, durable, and testable. Adjacent Skills must have explicit routing rules. Preserve deferred and rejected candidates with canonical reasons.

Orchestrate multiple agents safely

Read [references/multi-agent-orchestration.md](references/multi-agent-orchestration.md) before delegating research or validation. Use [assets/agent-task.template.json](assets/agent-task.template.json) for isolated assignments and require [assets/agent-result.template.json](assets/agent-result.template.json) outputs.

Parallel workers may extract claims; they must not write final recommendations or a shared ledger concurrently. The controller owns scope, identifiers, merge, conflict handling, gates, and final synthesis. Keep V3 blind evaluators isolated from expected answers and diagnoses.

When the host lacks workers, run the same task packets sequentially and mark execution_mode: sequential-fallback. Parallelism changes speed, not correctness requirements.

Validate artifacts and behavior

Read [references/evaluation-standard.md](references/evaluation-standard.md). Generate atomic test cases from [assets/evaluation-cases.template.json](assets/evaluation-cases.template.json). Initialize a pack with:

python3 scripts/init_pack.py  --industry "" --mode full

Validate static structure, evidence relations, candidate coverage, atomic test completeness, formulas, and parallel-run manifests when present:

Run [scripts/initpack.py](scripts/initpack.py) to initialize, [scripts/validatepack.py](scripts/validatepack.py) for pack checks, [scripts/runformulatests.py](scripts/runformulatests.py) for reproducible calculations, and [scripts/validateparallelrun.py](scripts/validateparallelrun.py) when an orchestration manifest exists.

python3 scripts/validate_pack.py 
python3 scripts/run_formula_tests.py 
python3 scripts/validate_parallel_run.py 

Use [scripts/mergeagentresults.py](scripts/mergeagentresults.py) only to combine isolated worker JSON into a review queue. It may deduplicate exact records but must not resolve substantive evidence conflicts automatically.

Static checks do not prove real triggering or execution. Prefer independent blind runs with the Skill artifact, realistic prompt, and raw inputs only. If unavailable, mark the result not-tested or self-test-fallback; never present it as independent validation.

Deliver the contract

Read [references/output-contract.md](references/output-contract.md) before creating or auditing a pack. Use [assets/executive-brief.template.md](assets/executive-brief.template.md) for the decision page and [assets/report-structure.template.md](assets/report-structure.template.md) for long-form synthesis. Read [references/platform-compatibility.md](references/platform-compatibility.md) before installing for Claude, Claude Code, ChatGPT, Codex, or another host.

Finish with:

• scope and evidence cutoff;
• decision findings with confidence and counterevidence;
• opportunity and entry recommendation, or the exact reason one is blocked;
• critical unknowns, stop rules, and next validation actions;
• capability coverage and all accepted/deferred/rejected candidates;
• V1/V2/V3 status and overall readiness;
• executed, manual, failed, and untested checks.

Read [examples/sample-requests.md](examples/sample-requests.md) only when designing trigger tests or diagnosing scope confusion. Use [tests.json](tests.json) when changing this meta Skill's trigger or workflow.

Stop conditions

Stop and request direction when the scope or geography remains decision-changing ambiguous; decisive private data is inaccessible; evidence is too weak for a high-consequence recommendation; credible sources conflict on a critical claim; or the user requests fabricated certainty, approval, credentials, or evidence.

Require explicit confirmation before publishing, messaging, purchasing, paying, deploying, deleting, overwriting, or making another consequential external change. For medical, legal, regulated financial, safety-critical, or controlled industries, provide research and decision support only and require current jurisdiction-specific professional verification before operational use.

License

Distribute and modify this Skill under the [MIT License](LICENSE.txt).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: shangdizhiyan
• Source: shangdizhiyan/Vic-distill-industry-skills
• License: MIT

Install and usage instructions live in the source repository linked above.