diff --git a/docs/specs/sf-prompt-modularization.md b/docs/specs/sf-prompt-modularization.md
new file mode 100644
index 000000000..5432c5f9a
--- /dev/null
+++ b/docs/specs/sf-prompt-modularization.md
@@ -0,0 +1,129 @@
+# SF Prompt Modularization
+
+Human-readable contract for the manifest-driven prompt context system.
+
+## Status
+
+M004 — Phase 3 complete. All builders migrated to `composeUnitContext` v2.
+
+## Overview
+
+Every autonomous unit type (plan-milestone, execute-task, deploy, …) declares
+what context it needs in a **manifest**. The composer reads the manifest and
+orchestrates artifact resolution, producing the `## Inlined Context` block that
+each prompt template receives.
+
+This replaces the pre-Phase-3 approach where each builder hand-wired file reads
+and DB queries inline.
+
+## Two composer APIs
+
+| API | Purpose | When to use |
+|-----|---------|-------------|
+| `composeInlinedContext(unitType, resolveArtifact)` | v1 — resolves `artifacts.inline` keys only | **Deprecated.** Retained in `unit-context-composer.js` for external consumers. Do not use in new builders. |
+| `composeUnitContext(unitType, opts)` | v2 — resolves `prepend`, `inline`, `excerpt`, and `computed` artifacts | **Use for all new builders.** All in-tree builders now call this. |
+
+### `composeUnitContext` contract
+
+```ts
+composeUnitContext(unitType: string, opts: {
+  base: string;                          // repo root
+  resolveArtifact?: (key: string) => Promise<string | null>;
+  resolveExcerpt?: (key: string) => Promise<string | null>;
+  computed?: Record<string, { build: (inputs, base) => string | null }>;
+}) => Promise<{ prepend: string; inline: string }>
+```
+
+- **prepend** — computed blocks that should appear before the main inline content.
+- **inline** — the joined context block suitable for `{{inlinedContext}}` substitution.
+- Unknown `unitType` returns `{ prepend: "", inline: "" }` — callers fall through to default behavior.
+- Missing resolvers or registry entries are skipped silently (no error).
+
+## Manifest schema
+
+```ts
+interface UnitManifest {
+  skills: { mode: "all" | "planning" | string[] };
+  knowledge: "full" | "scoped" | "critical-only" | "none";
+  memory: "prompt-relevant" | "critical-only" | "none";
+  codebaseMap: boolean;
+  preferences: "active-only" | "none";
+  tools: string[];
+  prepend?: string[];           // computed artifact ids (e.g. ["overrides"])
+  artifacts: {
+    inline: string[];           // static artifact keys, resolved in order
+    excerpt: string[];          // excerpt artifact keys
+    onDemand: string[];         // referenced but not auto-inlined
+    computed?: string[];        // async computed artifact ids (e.g. ["knowledge", "graph"])
+  };
+  maxSystemPromptChars: number;
+}
+```
+
+### Stable artifact keys (`ARTIFACT_KEYS`)
+
+Keys are **stable identifiers**, not paths. The resolver maps a key to its
+source (file, DB row, computed block).
+
+Milestone-scoped: `roadmap`, `milestone-context`, `milestone-summary`,
+`milestone-validation`, `milestone-research`, `milestone-plan`
+
+Slice-scoped: `slice-context`, `slice-research`, `slice-plan`, `slice-summary`,
+`slice-summaries`, `slice-uat`, `slice-assessment`
+
+Task-scoped: `task-plan`, `task-summary`, `prior-task-summaries`,
+`dependency-summaries`, `blocker-summaries`
+
+Project-scoped: `requirements`, `decisions`, `project`, `templates`, `queue`
+
+Validation-scoped: `verification-classes`, `outstanding-items`, `previous-validation`
+
+History-scoped: `prior-milestone-summary`
+
+## Migration decision matrix
+
+| Builder complexity | Recommendation | Example |
+|-------------------|------------------|---------|
+| Simple (1–3 static artifacts, no knowledge splice) | Full v2 via `composeUnitContext` | deploy, release, rollback, challenge |
+| Medium (4–6 static artifacts, optional knowledge/graph) | v2 inline + imperative knowledge splice | research-milestone, complete-slice, reassess-roadmap |
+| Complex (many artifacts, computed prepend, conditional logic) | v2 with full computed registry | plan-milestone, research-slice, plan-slice, replan-slice |
+| Fully imperative (no composer call) | Keep imperative; add manifest for documentation | execute-task, reactive-execute |
+
+## Budget and caching
+
+- `resolvePromptBudgets()` computes context-window-scaled budgets once per
+dispatch tick (1-second TTL). Do not add additional caches.
+- `capPreamble()` caps inlined context at `min(30K chars, scaled inline budget)`.
+- Exceeding `maxSystemPromptChars` logs a telemetry event; the composer does not
+truncate.
+
+## Verification
+
+- `auto-prompts-phase3.test.mjs` — v2 contracts for plan-milestone,
+  replan-slice, validate-milestone, research-slice.
+- `auto-prompts-complete-slice.test.mjs` — complete-slice closeout contract.
+- `auto-prompts-v2-migration.test.mjs` — regression guard for all migrated
+  builders (research-milestone, run-uat, deploy, smoke-production, release,
+  rollback, challenge).
+- `unit-context-manifest-computed.test.mjs` — manifest `computed` keys match
+  builder behavior.
+
+## Adding a new artifact key
+
+1. Add the key to `ARTIFACT_KEYS` in `unit-context-manifest.js`.
+2. Add the key to the `inline` / `excerpt` / `onDemand` / `computed` list of
+   every manifest that needs it.
+3. Implement the resolver in the builder's `resolveArtifact` function.
+4. Add a regression test that asserts the artifact appears in the prompt.
+
+## Adding a new computed artifact type
+
+1. Define the builder function (e.g. `inlineKnowledgeBudgeted`).
+2. Register it in the `computed` registry passed to `composeUnitContext`:
+   ```js
+   computed: {
+     knowledge: { build: (inputs, base) => inlineKnowledgeBudgeted(base, inputs.keywords) }
+   }
+   ```
+3. Add the id to `manifest.artifacts.computed`.
+4. Add a test that asserts the computed block appears in the output.