Pattern: Source grounding (cite official docs)

Ground every framework/library decision in official documentation: look it up, cite the source, and explicitly flag anything that could not be verified as unverified. Counters the model’s tendency to emit plausible-but-outdated API patterns from training data. The output carries its provenance, so a reviewer can check the citation rather than trust the assertion.

Applied by (backlinks)

Addy Osmani — Agent Skills:

• addy-source-driven-development — verify, cite, flag the unverified.

Compound Engineering:

• ce-ideate — grounds candidate directions in repo + web evidence.
• ce-pov — dual-grounded (repo + web) graded verdicts on external inputs.
• Research sub-agents — ce-web-researcher, ce-external-evidence-researcher, ce-slack-researcher, ce-best-practices-researcher, ce-framework-docs-researcher gather cited evidence.

gstack:

• gstack-office-hours · gstack-design-consultation · gstack-document-generate — research the landscape/codebase first (gstack’s “Search Before Building” ethos).

Seeds:

• seeds-plan-prompt — emits prior_art records ({id, type, summary, relevance}) mined from mulch alongside each plan section and instructs the model to ground its answer in them; the grounding corpus is the project’s own accumulated conventions, decisions, and failures rather than external documentation.

pstack — pointed at the organizational record rather than the public one, which is new here:

• pstack-why — discovers available MCP servers at run time and fans one investigator across each of seven evidence categories: source control, issue tracker, long-form docs, real-time chat, infrastructure observability, error tracking, and the product analytics warehouse. Ships per-source playbooks (Linear, Notion, Slack, Sentry, Datadog, Databricks, postmortems, code archaeology), skips an unavailable MCP and says so, and treats a null result as a finding. Seeds every investigator with a code anchor built inline from git blame, git log --follow, and gh pr view.
• pstack-recall — reuses that machinery with the question re-pointed from why was this built to what is the current state, what has been tried and did not hold, and what are users still reporting.

builder-skills:

• builder-read-the-damn-docs — signature. An explicit trigger list (latest/current/official, new or upgraded packages, fast-moving APIs, auth/billing/migrations, deprecation errors, and “you catch yourself about to write ‘usually’, ‘probably’, ‘I think’”), npm view <pkg> version before adding a dependency, and a stated fallback when docs cannot be reached.
• builder-visual-plan — “Research before you draft”, naming real files and symbols and leading with reuse.

See Also

• pattern-adversarial-review — the complementary “distrust the output” technique.
• stage-implement — where source-grounded code is written.