deep-research-discipline
Evidence-first workflow for multi-source deep research: requires cited sources, three-state confidence labels, cross-validation, and an independent review gate. Use when asked to research a topic in depth, verify claims across several sources, compare options, or produce a cited report. Tool-agnostic; needs only search, page fetch, and file read/write. Not for quick lookups, single-source summaries, or real-time data.
Deep Research Discipline
A process for producing research reports whose claims can be traced, graded, and challenged — rather than a search recipe.
This skill was developed and refined through agent operation, not hand-authored as a prompt template. See the repository README for provenance.
Applicability
Use for
- Multi-source fact gathering where being wrong is costly (market, technical, policy, competitor, standards).
- Claims that need to be separated into verified / single-source / unresolved.
- Reports that another person or agent must be able to audit later.
Do not use for
- A single lookup, a definition, or a quick link summary.
- Real-time values (prices, scores, incidents) where any cached source is stale by definition.
- Decisions requiring private/internal data the researcher cannot access.
- Legal, medical, or safety-critical advice — this process improves traceability, not professional judgment.
Capability boundary
Required (the workflow cannot run without these):
- Keyword search over the public web.
- Fetching a URL and extracting readable text.
- Reading and writing local files.
Optional (improves coverage; absence causes explicit degradation, never silent degradation):
- Date-range or recency filtering.
- PDF / Office / scanned-document extraction.
- Social or community sources (forums, social platforms).
- Code execution (for calculations over collected data).
Known limits — state these in the report when relevant:
- Paywalled, login-gated, or deleted content may be unreachable; search snippets are not evidence.
- Search indexes lag; a “not found” result does not prove non-existence.
- Non-English and offline sources are under-covered unless deliberately searched.
- Secondary reporting can inherit the errors of its primary source.
- This process reduces the probability of unsupported claims; it does not eliminate error.
Step 0 — Capability check
Before searching, list which required and optional capabilities are actually available in this environment, and write one line in the working notes:
capabilities: search=<yes/no> fetch=<yes/no> files=<yes/no> dates=<yes/no> documents=<yes/no> social=<yes/no>
Any capability marked no must appear later in the report's limitations section. Completion criterion: the capability line exists and matches what the tools actually do.
Step 1 — Frame the question
Split the request into 3–6 answerable sub-questions. For each, note the claim type (fact, number, date, causal explanation, opinion) and what evidence would settle it.
Completion criterion: every sub-question has a claim type and an acceptable evidence definition.
Step 2 — Search anchors
For each sub-question, derive anchors before searching: the exact entity name, the likely original publisher (regulator, vendor, standards body, paper), and a fallback query if the first returns noise.
Prefer the most primary source that can answer it: official document > dataset > first-hand report > reputable secondary > aggregator > social post.
Completion criterion: each sub-question has at least one named primary-source candidate.
Step 3 — Collect with independence
For every material claim, collect at least two sources that are independent — not two outlets quoting the same press release or the same dataset.
Record for each source: URL, publisher, publication date, retrieval time, and whether it is primary or secondary.
Completion criterion: every claim in the final report maps to at least one recorded source; material claims have two independent ones or an explicit single-source flag.
Step 4 — Grade the evidence
Label every material claim with one of three states:
| State | Meaning |
|---|---|
verified | Two or more independent sources agree, or one authoritative primary source states it explicitly. |
single-source | Only one source, or multiple sources derived from one origin. |
unresolved | Sources conflict, or no adequate source was found. |
Numbers, dates, and version identifiers must carry the source and retrieval date inline.
Completion criterion: no material claim is left unlabeled; conflicts are shown, not averaged away.
Step 5 — Resolve conflicts
When sources disagree, do not pick the more convenient one. Check, in order: publication date, whether one source cites the other, definition differences, scope differences, and known errata. If still unresolved, report both with the conflict stated.
Completion criterion: each conflict has either a resolution with evidence or an explicit unresolved entry.
Step 6 — Write the report
Use this structure:
# <topic>
## One-line answer
## Key findings (3–5, each with evidence state)
## Key data (number, source, date)
## Conflicting or missing evidence
## Analysis (per sub-question, traceable)
## Limitations (capability gaps, paywalls, coverage)
## Independent review record (BLOCKING)
## References (URL + retrieval time)
Completion criterion: every section present; the limitations section names every no capability from Step 0.
Step 7 — Independent review (blocking)
Have a separate reviewer — a different agent, session, or person — re-check 3–5 core data points against primary sources, without seeing the draft's conclusion first. Record each as PASS / FAIL / INCONCLUSIVE with the reviewer's own command or citation.
If no independent reviewer is available, say so explicitly in the report and mark the review section not performed. Do not present self-review as independent review.
Completion criterion: the review section exists, names the reviewer type, and lists per-point verdicts or an explicit not performed.
Degradation rules
- Missing social capability → “community sentiment not covered”, never implied.
- Missing document extraction → “original document not read; summary only”.
- Search returns nothing → run at least two alternative query formulations and one alternative source type before concluding absence.
- Never convert a snippet into a quotation; open the page or mark it unread.
Final self-check
- Every number has a source and a date.
- Every claim has a state label.
- Material claims have two independent sources, or are flagged.
- Limitations list matches the Step 0 capability line.
- Review section is real, or explicitly
not performed.
Optional channel mapping: see references/adapters.md.
- Downloads
- 189All time
- Security audit
- Pass
- Last updated
- 2w ago
- Current version
- v1.0.0
- License
- MIT-0
