# Intake > Bootstrap the curated context baseline for a brownfield (or near-greenfield) repo: scan the code and draft ALL curated docs with per-statement provenance markers, interview the developer through every gap/assumption/inconsistency, reconcile via the checkpoint loop, then seal as the initial curation baseline commit. Use when: (1) a repo has code but no curated docs/ baseline (docs/PRODUCT.md missing), (2) an explicit drift audit of the curated docs is requested, (3) a near-greenfield project needs its baseline before the first PROJ. Not for: per-PROJ documentation (use documentation), ongoing doc updates (P7 curation owns those), plan approval (use checkpoint). - Skill: `silviobeer/intake` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add silviobeer/intake` - Raw SKILL.md: https://api.skillmd.com/api/skills/silviobeer/intake/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security - Author: silviobeer (https://skillmd.com/u/silviobeer) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/silviobeer/intake --- # Intake — Bootstrap the Curated Context Baseline Owns the brownfield bootstrap of the agent workflow (CONCEPT.md §3). Produces the ONE curated docs baseline that every later phase injects into implementers — as early as possible, while the codebase is small and the interview is cheap. After the seal, P7 curation owns ALL updates; this skill is re-run only as a deliberate drift audit, never automatically. **The core principle: extraction and interview are two different sources, and this skill needs both.** Code tells what IS. Only the developer tells what is INTENDED — product scope and non-goals, security requirements, which of two inconsistent conventions is the rule going forward. This is a collaborative session, not a batch job. ## Codex Adaptation This skill is aligned with the Claude variant. In Codex: - Reusable skill assets live under `~/.codex/skills/...` instead of `~/.claude/skills/...`. - No AskUserQuestion tool and no Explore subagents: run the three scan passes sequentially yourself, and run the interview as numbered question blocks in the conversation (one topic per block, wait for the answers before the next block). - Everything else — provenance markers, the interview queue, the checkpoint reconcile, the seal check — is identical. ## Output (the baseline — all eight files) | File | Content | Cap | |---|---|---| | `docs/PRODUCT.md` | what the product is, for whom, non-goals | ½ page (≤30 non-blank lines) | | `docs/ARCHITECTURE.md` | the load-bearing architecture truths; details → `docs/architecture/` | ≤200 lines | | `docs/GUIDELINES.md` | conventions that ARE the rule (incl. Known Debt notes for losing patterns) | — | | `docs/DESIGN-SYSTEM.md` | design **rules** extracted from code — tokens, scales, patterns, do/don't — or a pointer to **frontend-design** (1c) if none exist; never invented. No component inventory (that's `components.md`) | ≤80 lines | | `docs/components.md` | component registry — **generated** by `scripts/gen-component-registry.mjs` from the doc block above each component export | — | | `docs/security-baseline.md` | auth model, secret handling, input validation floor | — | | `docs/test-conventions.md` | test runner, layout, naming, what must be tested | — | | `AGENTS.md` (root) | initial durable agent rules | ≤40 non-blank lines | `docs/DESIGN-SYSTEM.md` and `docs/components.md` are only required to seal when the repo has a component directory (`src/components` or `src/features/*/components`); backend-only repos skip both (see `scripts/intake-seal-check.sh`). **Existing AGENTS.md (or other curated docs) are baseline truth, never overwritten.** If a file already exists, the draft is a MERGE: every existing rule/statement is kept unless an interview decision explicitly retires it (recorded in `decisions.md`); new content is appended with provenance markers like everything else. The drift-audit variant works the same way — deltas, not rewrites. ## Workflow ### 0. Precondition - No curated baseline yet (`docs/PRODUCT.md` missing), OR the user explicitly asked for a drift audit of the existing baseline. - Working space: `specs/intake/` (drafts + decision log). This is pre-PROJ — there is **NO state.json**; state.json is born at CP1. - Drift audit variant: draft the delta against the existing docs instead of from scratch; everything else (interview, checkpoint, seal) is the same path. ### 1. Scan & draft (automated, provenance-marked) Run three read passes over the repo (spawn Explore agents where useful — they locate, you judge): - **Arch Hat:** entry points, module boundaries, data flow, persistence, integrations → ARCHITECTURE draft (+ `docs/architecture/` for detail). - **QA Hat:** test setup, error-handling patterns, validation, auth → test-conventions + security-baseline drafts. - **component-scout:** run `node scripts/gen-component-registry.mjs`, then write the missing doc blocks into the components it reports as undocumented (purpose + "not for" are interview material, not extractable) and regenerate → components.md registry; design tokens (tailwind config, CSS custom properties) → DESIGN-SYSTEM draft. If no real tokens exist, write a pointer to the **frontend-design** skill (1c) instead of inventing a design system. Draft ALL eight files into `specs/intake/` with a provenance marker on EVERY statement: - `[extracted: ]` — derived from code/README/git history, with the evidence reference - `[assumed]` — plausible inference that needs developer confirmation - `[gap: ]` — cannot be derived from code; must be asked While drafting, collect every detected **inconsistency** (e.g. "errors via Result type in ~60% of modules, thrown exceptions in ~40%") into the interview queue — never silently pick a winner. Near-greenfield variant: almost no code → extraction is thin and sourced from concept/PRDs; the interview carries the weight. Same files, same markers, same path. ### 2. Developer interview (point by point) Build ONE numbered queue: every `[gap]`, every `[assumed]`, every detected inconsistency. Work through it point by point as numbered question blocks — grouped by topic, never as one wall of questions. - Every answer becomes a GUIDELINES rule (or a PRODUCT/ARCHITECTURE statement, wherever it belongs). - For inconsistencies: the developer picks the rule going forward; the LOSING pattern becomes a **Known Debt** note in GUIDELINES — recorded, not silently rewritten. - `[assumed]` items the developer confirms lose their marker; corrected ones are rewritten from the answer. ### 3. Reconcile via checkpoint (4a, Bootstrap Variant) Route to the **checkpoint** skill's Bootstrap Variant: the generated docs are HYPOTHESES, not truth. Same point-by-point pattern — every review point ends in exactly one of adopt / change / reject / defer, recorded in `specs/intake/decisions.md` (`D-BOOTSTRAP-` ids, `templates/decisions.md.tmpl` frame) BEFORE the next point. Cascade every decision into the affected draft(s). ### 4. Seal 1. Move the reconciled drafts to their final locations (`docs/*`, root `AGENTS.md`) and STRIP all provenance markers — they are working syntax, the sealed baseline never carries them. 2. `bash scripts/intake-seal-check.sh` (copy from `~/.codex/skills/0b_intake/scripts/` together with `curation-caps.sh` if missing). It verifies: all eight files exist, zero residual markers, `specs/intake/decisions.md` carries `D-BOOTSTRAP` reconcile decisions (undiscussed drafts never seal), file caps hold. Red → fix, re-run; NEVER seal past a red check. 3. Commit: `docs: intake baseline — initial curated docs` (include `specs/intake/decisions.md`). From here on, P7 curation owns every docs/ update. → NEXT ACTION: start the first PROJ through the chain (**chain-guide** routes it); at P0 the setup skill compiles this baseline into the context bundles. ## Failure Behavior - The interview cannot proceed (developer unavailable) → leave the marked drafts in `specs/intake/`, write one line to `specs/intake/` README noting the open queue — the drafts are resumable; nothing is half-sealed into `docs/`. - `intake-seal-check.sh` red → the baseline stays uncommitted; fix and re-run. The check is idempotent.