Skill 3 Initiate Synthesis

Initiate Synthesis

Establish the Contract

If the mode is unclear, ask exactly:

repo, paper writing, or are we synthesising other types of sources of truth/data types today?

Do not ask when the request already establishes the mode. Infer what is safe and ask only for missing choices that would materially change the result.

Before substantive synthesis, establish:

• the question or decision the synthesis must answer;
• the intended artifact and audience;
• the source boundary, date boundary, and relevant exclusions;
• the appropriate rigor level: rapid, bounded, or systematic.

Use rapid for orientation from a convenience corpus, bounded for a defined supplied corpus, and systematic only when search, screening, extraction, and reporting can be made reproducible. Never call a convenience sample systematic or comprehensive.

Select one dominant mode:

• repo: code, configuration, issues, logs, documentation, runtime evidence, or deployment targets dominate.
• paper: research literature, drafts, figures, references, reviews, methods, results, or datasets dominate.
• knowledge: mixed documents, notes, URLs, datasets, transcripts, or a topic-level analysis dominate.

Preserve the inventory if the mode changes.

Build an Auditable Source Inventory

Inventory sources before drawing conclusions. Assign stable IDs (S001, S002, ...). For each source, record as available:

• path or canonical URL;
• source type, authorship/owner, date/version, and access date;
• origin: user, local, remote, or generated;
• access status and extraction status;
• duplicate, superseded, retracted, corrected, or overlapping status;
• inclusion decision and reason;
• likely relevance and important limitations.

Distinguish three counts: sources discovered, sources screened, and sources substantively examined. Report unreadable or unavailable sources. Do not imply coverage of files that were only listed or skimmed.

For a large corpus, use transparent staged sampling:

1. map file types, dates, topics, and obvious duplicates;
2. inspect a deliberately varied sample rather than only convenient or prominent items;
3. state the sampling rule and provisional findings;
4. expand until the requested coverage is reached or further sampling stops changing the synthesis materially;
5. record what remains unexamined.

Prefer primary material for claims about results, implementations, or policies. Use reviews and secondary sources for orientation and citation chaining, not as automatic substitutes for primary evidence.

Separate Observation, Interpretation, and Claim

Maintain a claim-evidence ledger for non-trivial conclusions. Use stable claim IDs (C001, C002, ...). Record:

• the precise claim;
• claim type: descriptive, comparative, causal, interpretive, predictive, or normative;
• supporting source IDs and exact locator (page, section, table, figure, line, commit, or timestamp);
• contradicting or qualifying source IDs;
• population/context and boundary conditions;
• appraisal concerns;
• status: supported, mixed, tentative, unsupported, or not assessed;
• provenance: source-explicit, user-stated, agent-inferred, or agent-generated.

Never upgrade an inference into a source-reported fact. Label calculations, transformations, classifications, and interpretations as derived. Preserve the chain from output claim to source passage or raw evidence.

Apply Critical Appraisal

Appraise evidence relative to the claim and study/source type; do not treat a single generic score as objective truth. Read references/synthesis-rigor.md when producing a literature review, evidence synthesis, high-stakes analysis, or any artifact presented as rigorous.

At minimum examine:

• relevance and directness to the question;
• design fitness for the claim type;
• selection, measurement, confounding, reporting, and sponsorship risks;
• sample/context adequacy and transferability;
• analytic transparency and reproducibility;
• consistency with other evidence;
• version currency and source independence.

Peer review, citation count, venue prestige, and author reputation are context, not validity guarantees. Multiple papers using the same dataset or research group may not be independent evidence.

Actively seek disconfirming evidence and alternative explanations. Retain null, negative, and inconvenient findings. Distinguish evidence of no effect from absence of evidence. Do not infer an effect from the number of positive studies alone; account for study quality, precision, dependence, and heterogeneity.

Synthesize by Mode

Repo Mode

Inspect the repository before editing. Identify stack, entry points, conventions, tests, deployment targets, and current worktree state. Treat code behavior, tests, logs, documentation, issue reports, and user statements as different evidence types; reconcile conflicts explicitly.

For diagnostic or comparative conclusions, link each conclusion to files, lines, commands, logs, or runtime observations. State whether behavior was inspected, tested, reproduced, or inferred. Preserve user changes and validate implementation outputs proportionately to risk.

Paper Mode

Define the target artifact and review question. For literature synthesis, make eligibility criteria operational before screening where feasible. For new searches, record databases/sites, exact queries, filters, dates, and citation-chaining steps. Keep discovery separate from eligibility decisions.

Extract comparable fields into an evidence matrix before drafting conclusions. Include study context, design, sample/data, intervention or phenomenon, comparator, outcomes, analysis, key findings, limitations, and appraisal. Adapt fields to qualitative, quantitative, mixed-methods, theoretical, or technical work.

Choose a synthesis method that fits the evidence:

• narrative synthesis for heterogeneous evidence with structured comparison;
• thematic or framework synthesis for qualitative concepts, with coding decisions visible;
• evidence mapping for coverage, clusters, and gaps without effect claims;
• quantitative pooling only when estimands, designs, and data are sufficiently compatible.

Explain heterogeneity rather than averaging it away. Compare contexts, constructs, methods, outcomes, and time horizons. Draft from the evidence matrix and claim ledger, not memory. Verify every bibliographic identity and ensure cited sources entail the nearby statement. Never invent citations, metadata, quotations, page numbers, or findings.

Knowledge Mode

Define the unit of analysis and classification scheme before bulk classification. Separate source-authored categories from agent-generated categories. Test categories on a varied sample, record ambiguous or multi-label cases, and revise the scheme transparently.

For comparisons, use common dimensions and disclose missing data. For topic discussion without a requested artifact, remain concise but preserve epistemic labels and source grounding; do not use an arbitrary character limit when nuance is needed.

Produce the Artifact

Unless the user requests another form, structure a substantive synthesis as:

1. question, scope, and rigor level;
2. corpus and method, including coverage and exclusions;
3. findings organized by claims or themes;
4. contradictory, null, and boundary evidence;
5. confidence and limitations;
6. gaps and defensible next steps;
7. source inventory or citation list and, when useful, claim-evidence ledger.

Calibrate language to evidence. Reserve causal language for designs that support causal inference. Use concrete qualifiers such as population, setting, outcome, and timeframe instead of vague hedges.

Before delivery, audit:

• every material claim is traceable;
• locators and citation identities were verified;
• conclusions do not exceed source scope;
• contradictions and negative evidence are visible;
• generated interpretation is labeled;
• discovered, screened, included, and unread counts are not conflated;
• limitations describe consequences, not boilerplate;
• the artifact distinguishes what is known, uncertain, disputed, and missing.

If evidence is insufficient, say so and specify what would resolve the uncertainty. Do not manufacture consensus or completeness.

Maintain Session State

Keep a compact working map of the contract, source inventory, eligibility decisions, claim ledger, assumptions, and unresolved questions. Update it as sources or decisions change. For multi-session or high-rigor work, persist these as artifacts rather than relying only on chat history. Use the templates in references/synthesis-rigor.md.