---
name: cp-skill-write-multistage
description: Write or rebuild a KB artifact through reconstruction, claim disposition, drafting, audit, and promotion. Use when claims need grounding, synthesis, or separation across multiple artifacts; avoid it for settled local edits.
type: kb/types/instruction.md
user-invocable: true
allowed-tools: Read, Edit, Write, Grep, Glob, Bash, Skill, Task
argument-hint: "[target path | collection/type/topic] [source paths or brief]"
context: fork
---

# cp-skill-write-multistage

## EXECUTE NOW

**Target and inputs: $ARGUMENTS**

Develop one substantive KB artifact through independent reconstruction, claim disposition, a claim skeleton, drafting, audit, and reconciliation. Keep every intermediate artifact under one `kb/work/multistage/` workshop. Do not add workflow-state fields to the target artifact's frontmatter, and do not write the target until promotion.

This workflow requires fresh sub-agent contexts. If the runtime cannot create them, initialize the workshop, record the limitation, and stop before source reconstruction. Do not imitate source-first independence in a context that has already read the incumbent draft.

## Step 1 - Resolve The Target And Contract

Determine whether this is:

- **Edit mode:** `$ARGUMENTS` identifies one existing Markdown artifact.
- **New-write mode:** `$ARGUMENTS` identifies a collection, type, topic, or intended path.

**Conditional external-literature disposition.** If the task explicitly asks
whether an existing claim-bearing artifact duplicates, restates, or is
subsumed by external literature, or asks whether to keep, rewrite, thin, merge,
retire, or remove it from a cohort on that basis, load the framework procedure
`assess-a-claim-bearing-artifact-against-external-literature.md` before
continuing. In an installed project, read it from
`kb/commonplace/instructions/`; in the Commonplace source checkout, read it
from `kb/instructions/`. Treat that procedure as the governing source-selection
and disposition contract, add its required assessment records to this skill's
workshop checklist, and use this skill's later stages to realize any selected
artifact change. Do not load it for ordinary named-source grounding, synthesis,
or local revision.

Resolve the target collection to a directory under `kb/` with a local `COLLECTION.md`, and read that file in full.

In edit mode, read `type:` from the incumbent frontmatter and open that type specification. If the file has frontmatter but no `type:`, stop and repair that structural problem first. If it has no frontmatter, treat it as implicit `text`; do not invent a type or type specification.

In new-write mode, default an unspecified collection and type to `kb/notes/` and `kb/types/note.md`. If the user or calling workflow supplied a type path, open it and verify from its own frontmatter that it is a type spec. For a shorthand type name, search Markdown files under `kb/types/` and every collection `types/` directory below `kb/`, inspect their own opening frontmatter, and require exactly one type-spec doc with that `name:`. If none or several match, stop and report the matching paths; do not guess or apply collection-specific precedence. Type lookup identifies the contract rather than collection eligibility, which `commonplace-validate` owns at promotion. Do not add a `kb/work/` branch. An explicit request for `text` means frontmatter-free Markdown rather than a type path.

In new-write mode, run one targeted near-duplicate search using distinctive title or topic terms. Prefer revising a near-duplicate over creating another artifact. Derive a provisional lowercase-hyphenated target path with a filename of at most 70 characters from the requested title or topic before creating the workshop. Use this initial path as the immutable run key. Treat a user-supplied path as fixed unless it violates the collection or type contract; otherwise, if the final title changes the destination, update the current target in `README.md` without changing the run key.

In edit mode, read the incumbent in full, run one backlinks lookup, and preserve a copy as `original.md` in the workshop. Remove `user-verified` from the eventual candidate after any substantive edit unless the user explicitly re-verifies it.

When the task is a mechanical update, a local prose edit, or a straightforward write whose claims, evidence, and structure are already settled, stop and explain why the multistage path is unnecessary. Ask whether the user wants to continue with `cp-skill-write`, and invoke it only after explicit confirmation. When no library artifact is yet intended, explain that the task belongs in an exploratory workshop and ask before creating one.

## Step 2 - Create Or Resume The Workshop

Search `kb/work/multistage/` for an unfinished multistage workshop whose declared immutable run key or current intended target path exactly matches this run; do not match paths mentioned only in pending handoffs or prose. If exactly one exists, resume it. If several exist, stop and ask which one to use. Otherwise create:

```text
kb/work/multistage/multistage-write-<short-topic>-<YYYYMMDD>/
```

If that directory already exists for another target, append the smallest available numeric suffix, beginning with `-2`.

Create `README.md` with:

- the immutable run key, current intended target path, mode, collection, and type;
- the workshop's source/input paths;
- a checklist for `brief.md`, `reconstruction.md`, `claim-disposition.md`, `claim-skeleton.md`, `draft.md`, `audit.md`, `candidate.md`, conditional `acceptance.md`, and promotion;
- unresolved human decisions and blockers;
- pending handoffs, each with its target, proposed delta, user authorization, and order when known;
- whether acceptance review is required, not required, or complete.

The checklist and stage files are the workflow state. Do not introduce a `stage` frontmatter field. Mark a stage complete only after its file is non-empty, contains the required items for that step, and has no blocker that the next step would hide. If an upstream artifact changes, uncheck and regenerate every dependent stage before promotion.

Add a one-line entry for the active run to `kb/work/README.md`. Preserve unrelated edits in that file; if an overlapping uncommitted change makes the update unsafe, record the pending index update in the workshop `README.md` and report it rather than overwriting another agent's work.

## Step 3 - Write The Brief

Write `brief.md` before delegating any prose. Include only information fixed by the task:

- the question or decision the artifact must address;
- intended audience and what the reader should understand, infer, or do;
- intended target path, mode, collection, and type;
- target claim or purpose supplied by the user, without expanding it;
- scope, exclusions, required terminology, and collection/type constraints;
- source and evidence paths available to the run;
- user directions and retained intent supplied to the run, with each memory input's source, subject, scope, and authoritative or advisory role, distinguishing inputs that can select intent from evidence that can warrant claims;
- known uncertainties, missing evidence, and decisions reserved for the user.

Repository and collection contracts may supply the acceptable contribution class, quality bar, and a default audience. They do not by themselves select the artifact's governing question, claim, or purpose. Carry choices already fixed by the task, incumbent artifact, or supplied retained intent into the brief without asking the user to restate them. Current user direction prevails. If retained intent conflicts with the incumbent or another applicable input and no explicit precedence resolves the conflict, leave the choice unresolved. Do not treat remembered intent as meaning extracted from the bare request, model prior, or factual warrant. This skill consumes memory supplied through the retained-intent input but does not search raw interaction history itself.

If several materially different contributions still fit, record `DECISION NEEDED: intended contribution (specification gap)` in `brief.md` and the workshop `README.md`, then stop before reconstruction. A stronger model may use supplied context better, but greater capability does not make one of several compatible commissions authoritative. Source reconstruction must not choose the commission.

Acquire or ingest every named input needed to answer the governing question before continuing. Do not use search snippets as evidence. Missing intent is **blocking** when it leaves materially different commissions open. Missing evidence is **blocking** when the artifact cannot answer its governing question without asserting the missing claim. Pause at this step for either kind of blocker. A gap is **non-blocking** when the claim can be omitted or the uncertainty can honestly remain part of the final artifact; record it for reconstruction.

## Step 4 - Reconstruct From Sources In A Fresh Context

Launch one fresh, single-use sub-agent. Give it only `brief.md` and the exact source/evidence paths listed there. Do not give it `original.md`, any prior draft, or conclusions from another reviewer. Tell it not to search for or read those files.

Have it write `reconstruction.md` containing:

- the material facts, mechanisms, distinctions, quantities, and definitions supported by the inputs;
- the source or user direction supporting each material item;
- conflicts among inputs and differences in evidential strength;
- inferences stated as inferences rather than source facts;
- unresolved questions and explicit `EVIDENCE NEEDED`, `DEFINE`, or `DECISION NEEDED` markers where appropriate;
- details that are available but irrelevant to the target question, so they are not reintroduced merely because they are concrete.

Keep the reconstruction proportional to the inputs. Do not repeat the same limitation under several headings, derive unrequested statistics, or enumerate unavailable details that do not affect the target claim. Do not ask for polished prose. The reconstruction is an independent account against which later prose can be audited.

## Step 5 - Dispose Candidate Claims Before Choosing The Artifact Shape

Launch a new single-use claim architect with `brief.md`, `reconstruction.md`, and the target collection/type contracts. Do not initially give it `original.md` or any draft. It may run targeted title and description searches and open plausible existing notes for each candidate claim discovered during reconstruction. In edit mode, give it the current target path as an exclusion: it must not open that file even when a search surfaces it.

Have it write a `## Source-first disposition` section in `claim-disposition.md`. Inventory every candidate durable claim needed to answer the governing question and record:

- the claim in one sentence;
- its relation to the user-supplied target;
- its evidential basis;
- whether an existing artifact already states it adequately;
- one disposition: `central contribution`, `cite existing`, `fold into existing`, `separate new artifact`, `support/example/scope only`, or `omit/retain in workshop`;
- the existing or proposed target path when applicable;
- why the disposition preserves a useful citation and revision boundary.

In edit mode, wait until the source-first section is saved before giving the same architect `original.md`. Then have it append `## Incumbent reconciliation` without rewriting the source-first section. It must inventory every material incumbent commitment that the source-first pass omitted, merge duplicates explicitly, and give each remaining commitment the same disposition fields. The incumbent can reveal a claim that needs evaluation, but it is not evidence for that claim. If retaining or revising an incumbent commitment requires support absent from `reconstruction.md`, mark `EVIDENCE NEEDED` and return to Step 4 after acquiring the exact evidence path and adding it to `brief.md`. An unsupported commitment may be explicitly omitted only when the governing question and supplied intent do not require it; otherwise treat the missing evidence as blocking under Step 3. The architect must not open the live target or any draft during this reconciliation.

For a claim-bearing artifact, default to one atomic central contribution: one proposition another artifact can cite as a premise without inheriting an independent claim cluster. Evidence, mechanism, consequences, examples, and scope may remain when they establish, apply, or bound that proposition. A section that could be removed while leaving the central argument intact, and that another artifact may need to cite independently, is a separate claim rather than supporting completeness.

Do not use the `synthesis` trait merely because reconstruction produced several relevant claims. Use it only when the composition or inferential relation among already-citable components is itself the central contribution and no newly introduced component needs an independent citation or revision boundary. Definitions, specifications, articles, instructions, and other types whose contracts require multiple commitments keep their type-appropriate shape, but still dispose independent transferable claims instead of hiding them inside the artifact.

Apply these decision gates:

- If the task and evidence do not determine which of several claims is central, record `DECISION NEEDED: central contribution`, update the workshop `README.md`, and ask the user.
- If a discovered claim should substantively change an existing artifact other than the current target, record the proposed target and delta, then ask the user before folding it. Merely citing an adequate existing claim does not require confirmation.
- If more than one independent new artifact is warranted, record each proposed artifact and ask the user which to produce first and whether to create separate runs for the others. If the user already authorized the set and order, record that direction and continue with only the current artifact; every additional artifact gets its own run.
- Do not ask about claims classified as support, example, scope, omission, or adequate existing premises.

After a user decision, clear the corresponding decision marker in the workshop `README.md` and add the direction to `brief.md`. Before invalidating `claim-disposition.md`, copy every authorized but non-current fold or additional artifact into the README's pending handoffs.

If the decision changes target identity, mode, collection, or type—not merely the provisional filename of a new artifact within the same collection and type—return to Step 1 first. Re-resolve the target and contracts and replace target-specific incumbent and backlink inputs. Synchronize the target path, mode, collection, type, and source paths in `brief.md` and the workshop `README.md`, and replace the collection/type constraints in `brief.md`. If the selected target is an additional artifact rather than a replacement for the current one, leave it as a Step 10 handoff instead of retargeting this run.

Because the brief or target inputs changed, uncheck reconstruction and every dependent stage, then resume at Step 4. Do not regenerate only `claim-disposition.md`. Continue only when the rebuilt disposition names exactly one current central contribution or one type-appropriate practical purpose with no unresolved decision marker.

## Step 6 - Build The Claim Skeleton In A Fresh Context

Launch a new single-use sub-agent with `brief.md`, `reconstruction.md`, and `claim-disposition.md`. It may read a named source only to resolve an explicit reconstruction ambiguity. It must not read `original.md` or any draft.

Have it write `claim-skeleton.md` as a compact ordered plan containing:

- the one central claim or type-appropriate practical purpose fixed by `claim-disposition.md`;
- the work each section or paragraph must perform;
- each material assertion, its scope and confidence, and its evidential basis;
- the inferential links needed to move from evidence to conclusion;
- definitions or comparisons needed for truth conditions;
- unresolved markers, each classified as blocking, omittable, or suitable for an explicit published limitation or open question;
- tempting but irrelevant branches to omit.

Every planned paragraph must change what the reader understands, infers, or can do. Do not add setup, summary, or praise merely to make the artifact sound complete.

Do not proceed while a blocking marker remains. For each non-blocking marker, either omit the dependent claim or explicitly authorize its conversion into a published uncertainty, limitation, or open question. Update the workshop checklist and resume from reconstruction when new evidence changes the skeleton.

## Step 7 - Draft In A Fresh Context

Launch a new single-use writer with `brief.md`, `reconstruction.md`, `claim-disposition.md`, `claim-skeleton.md`, and the target collection/type contracts. Do not give it `original.md`.

Have it write `draft.md`. Require it to:

- realize the skeleton rather than discover new claims through fluent prose;
- preserve qualifiers, scope, uncertainty, and source distinctions;
- name mechanisms, comparison bases, and applicable scope where they affect the claim;
- express authorized uncertainty in reader-facing prose, but never copy workshop markers such as `EVIDENCE NEEDED` or `DECISION NEEDED` into the draft;
- use the simplest structure and language that carry the argument;
- omit frontmatter fields that the collection/type contract does not authorize.

The writer must not silently introduce a new commitment. When the prose appears to need one, insert `NEW COMMITMENT FOR AUDIT:` with the proposed claim instead of treating it as established.

## Step 8 - Audit Commitments Before Polishing

Launch a new single-use auditor. Give it `brief.md`, `reconstruction.md`, `claim-disposition.md`, `claim-skeleton.md`, `draft.md`, the relevant contracts and sources, and `original.md` in edit mode.

Have it write `audit.md` with anchored findings. Each finding begins with `Status: open` and recommends one action: `keep`, `remove`, `ground`, `clarify`, or `ask user`. `Keep` means that the cited draft text is already justified; the finding must name that basis. Audit in this order:

1. **Claim delta:** identify every draft commitment absent from the skeleton, every planned commitment omitted or altered, and every change in causality, scope, confidence, quantity, or recommendation. In edit mode, also verify that every material incumbent commitment appears in the incumbent reconciliation, then identify incumbent commitments the draft drops or changes.
2. **Artifact shape:** check every independent claim against `claim-disposition.md`. For a claim-bearing target, verify that the title, description, opening, and body expose one importable central proposition; flag any second cluster that should be cited, revised, folded, or promoted independently. Reject a `synthesis` trait used only to waive extraction.
3. **Grounding:** check material assertions against the reconstruction and sources. Distinguish source fact, user direction, inference, and unsupported completion.
4. **Specificity:** flag ambiguity that changes truth conditions, support, or implications. Ask for mechanism, comparison basis, or scope only where it is load-bearing; do not demand decorative detail.
5. **Relevance and audience:** flag undefined dependencies, displaced implications, irrelevant facts, and paragraphs that do no new work.
6. **Compression and prose:** only after the content audit, identify repetition, filler, unnecessary framing, and sentences whose syntax hides the claim.

Do not rewrite the draft in the audit. If a correct recommendation requires evidence or intent not present in the inputs, use `ask user` rather than guessing.

## Step 9 - Reconcile Into A Candidate

The orchestrating agent reads all workshop artifacts and writes `candidate.md` as a complete target artifact, including valid frontmatter when the selected type uses it. Preserve frontmatter-free implicit `text` as frontmatter-free text.

Resolve every audit finding explicitly in `audit.md`: add `Status: resolved` and a `Resolution:` naming the candidate change or the reason the text was kept. Use `Status: blocked` when evidence or a user decision is still missing. Do not promote while any finding remains open or blocked.

When reconciliation introduces material evidence not covered by `reconstruction.md`, return to Step 4. When it changes the central contribution, any claim disposition, or the justification for synthesis without new evidence, return to Step 5. When it changes only ordering or expression, continue.

For public-facing, high-stakes, causal, or quantitative work—or whenever the audit found material drift—launch one final fresh acceptance reviewer. Give it `brief.md`, `reconstruction.md`, `claim-disposition.md`, `claim-skeleton.md`, `audit.md`, `candidate.md`, the relevant contracts and sources, and `original.md` in edit mode. Have it write `acceptance.md` with `Verdict: PASS` or `Verdict: BLOCK` followed by anchored blockers only. Reconcile a block and rerun once, replacing `acceptance.md` with the current verdict; if material disagreement remains, ask the user. When review is not required, mark it `not required` in `README.md`.

## Step 10 - Promote And Validate

Promote only when:

- all workshop blockers and unintended unresolved markers are gone;
- `claim-disposition.md` names one current contribution, all independent claims have explicit dispositions, and every required user decision is recorded;
- each material commitment has an authorized basis, with inferences, uncertainties, and open work labeled appropriately;
- the candidate follows the collection and type contracts;
- every audit finding is resolved;
- any required acceptance review passes.

Before promotion, inspect `candidate.md` for every addition or material change
that depends on a named external source, regardless of the target collection.
This includes a source, URL, or ingest supplied as support; an added or changed
attribution, quotation, empirical result, or borrowed mechanism tied to a named
source; and a review finding that asks for exact claim/source grounding. A
passing mention or adjacent example not used as support is not a dependency.
In edit mode, compare against `original.md` and do not retrigger the guard for
unchanged source-dependent wording.

For each guarded dependency, resolve exactly one direct tracked
`kb/sources/<slug>.ingest.md` from the supplied ingest, canonical source URL, or
unambiguous source identity. Read its complete Quotes section and the
`semantic/grounding-alignment` gate from the installed framework gate catalog.

- When the retained verbatim quotes contain enough source material for the
  gate to judge the candidate's use, apply the gate directly to that use. Link
  the ingest without a snapshot marker, ignore every ingest section outside
  Quotes as source support, and keep target-specific transfer reasoning in the
  target.
- Use the snapshot route only when an earlier grounding run returned `snapshot
  required`. Put the exact marker `(snapshot required)` in the ingest link
  text. Derive the exact name-paired snapshot, require its exact-byte SHA-256
  and canonical source to match the ingest, read it, and apply the same gate to
  the candidate's use. Stop if the snapshot is absent, mismatched, or does not
  support the use.
- Otherwise, if Quotes is insufficient and the ingest exists, invoke
  `cp-skill-ground` with `Target: <ingest path>` and `Claim needed: <the
  source-side proposition or question>`, then act on its route: `quotes
  sufficient` or `quotes added` — re-read the Quotes section and apply the gate
  as in the first bullet; `snapshot required` — take the snapshot route above;
  a blocker — add it to the workshop `README.md` with the exact dependency and
  retain `candidate.md` and the workshop without changing the live target.
  Record every `quotes added` result, with the ingest path, in the workshop
  `README.md` and the final report so the append is never a silent side effect
  of a write.
- If the dependency names a URL with no tracked ingest, add a blocker naming
  the exact URL for a separate `cp-skill-ingest` run; this writer does not
  create source records.

If neither an exact ingest nor a canonical URL can be resolved, record that
source-identity blocker and ask for the missing identity. This writer invokes
the grounding skill but never edits an ingest itself, never creates one, and
introduces no separate result protocol. It reads a source snapshot only for a
declared `snapshot required` dependency. Do not begin any promotion write while a
source-dependency blocker remains. If applying the source gate or changing the
link materially changes the audited candidate, return to Step 9 and renew any
affected audit or acceptance work before promotion.

Identify each focused local source whose collection authorizes a source-to-target lineage footer. Validate it in its current state, and preserve a workshop copy of every source that will change. If a source is already invalid, stop before promotion.

Before writing, verify the candidate's frontmatter when applicable, required sections, and relative links as they will resolve from the target directory. Write `candidate.md` to the resolved target path. Preserve valid incumbent metadata and links unless the revision requires changing them. For a new artifact, derive a lowercase hyphenated filename of at most 70 characters from its title unless the user supplied a path. Never grant `user-verified` implicitly.

Run:

```bash
commonplace-validate path/to/target.md
```

Fix validation failures immediately. If they cannot be fixed within the established claim and contract, restore `original.md` byte-for-byte in edit mode or remove the newly created target in new-write mode, retain the workshop, and report the blocker.

After the target validates, add the authorized lineage footers and validate every changed source. Do not add a target-to-source lineage footer merely for symmetry. If a lineage edit cannot be made valid, restore every changed source and the target to their pre-promotion state, retain the workshop, and report the blocker.

After successful validation, remove the exact completed workshop directory and its `kb/work/README.md` entry unless the user asked to inspect or retain the run as an experiment or audit record, or any recorded user decision remains unexecuted — such as a confirmed fold into another artifact or an authorized additional artifact awaiting its own run. If retained, mark its state in `README.md` and keep its index entry. Treat such pending work as a handoff: report each fold or additional artifact and let the user decide what to do next; do not start another run automatically. When the user directs that a handoff has been declined or completed, mark it resolved and remove this workshop once no retention reason remains. Report what was removed or retained. Suggest `cp-skill-connect` for broader graph discovery.

## Verify

- The target remained untouched until a reconciled candidate was ready.
- Source reconstruction occurred in a fresh context that never saw the incumbent or draft.
- In edit mode, source-first disposition was saved before the incumbent was revealed, and every material incumbent commitment was then reconciled without treating it as evidence.
- Claim disposition preceded the skeleton; existing claims were cited or proposed for confirmed folding, and additional new artifacts were separated into their own user-authorized runs.
- A direction added to the brief after a user decision caused reconstruction and every dependent stage to be rebuilt.
- A claim-bearing target exposes one importable central proposition unless the disposition establishes that an irreducible synthesis is itself the contribution.
- The skeleton preceded prose and every material draft addition was audited.
- Missing knowledge remained visible instead of becoming plausible filler.
- Content review preceded compression and sentence polish.
- Workflow state lives only in the workshop, not in library frontmatter.
- The promoted artifact passes deterministic validation.