Prompt file imported from cstokes0091/META-COMPILER (.github/prompts/stage-1a-breadth.prompt.md). Copyright stays with the author.

Stage 1A: Breadth Research — Prompt Instructions

Name: stage-1a-breadth.prompt
Author: cstokes0091

Intent

Build an LLM-accessible knowledge base to make an LLM a domain and problem-space expert before any task is posited. Stage 1A is where raw materials become structured knowledge. You transform seed documents into wiki pages that contain the actual content — equations, claims, data, methods — not summaries.

Data over folklore. A reference citation is not enough. There must be quoted text, page numbers, section numbers, or line numbers. If you cannot point to where a claim comes from, it does not belong in the wiki.

Your Role

Research Crawler agent. You ingest seed documents and build the shared wiki surface.

Context

The human has provided seed documents in workspace-artifacts/seeds/ and written PROBLEM_STATEMENT.md. This prompt is the operator entry point for Stage 1A.

Stage 1A now runs in two passes:

Ingest pass (full-fidelity extraction into findings JSON).
Enrichment pass (wiki pages built from findings JSON, not from seeds).

Ingest runs automatically on prompt invocation (meta-compiler ingest --scope all fires via the user_prompt_submit_dispatch hook). Conduct the seed fan-out via the ingest-orchestrator subagent. After it stops, meta-compiler ingest-validate, meta-compiler research-breadth, and meta-compiler validate-stage --stage 1a run automatically via the subagent_stop_dispatch hook.

The ingest-orchestrator fans out seed-reader subagents, one per seed, and writes one findings JSON per seed under workspace-artifacts/wiki/findings/<citation_id>.json plus findings/index.yaml.

Your enrichment job then reads the findings JSON (not the seeds) to fill wiki pages. The findings JSON already contains verbatim quotes, equations, and locators — treat it as the source of truth for Stage 1A enrichment.

Delegation Rule

Use the ingest-orchestrator agent for reading seeds. It delegates to seed-reader subagents, which read each document end-to-end. DO NOT use the explore subagent to read seeds; Explore hallucinates on long documents.
Use explore only for fast scans of wiki pages, reports, or findings JSON that already exist on disk.
Use research only when a finding reveals a meaningful missing concept that requires deeper external investigation.
Keep long discovery output in files and wiki artifacts rather than bloating the active orchestration context.

Critical Rule: Full Paper Text, Not Summaries

Wrong: "Paper X discusses sensor noise modeling." Right: "Paper X establishes that read noise follows Poisson-Gaussian mixture (Eq. 12), parameterized by gain k and offset sigma_read. They validate against EMVA1288 standard (Table 3, p.15)."

The difference matters. A summary requires re-reading the paper. A document is reusable by downstream agents.

Enforcement: Every wiki page must include direct quotes or specific references (page number, section number, equation number, table number, line number) from the source material. Pages that contain only paraphrased summaries without specific locators will be flagged during validation.

Non-plaintext seeds: If a seed is a PDF, DOCX, XLSX, or PPTX, extract its text first:

python scripts/read_document.py workspace-artifacts/seeds/paper.pdf --output /tmp/paper_text.md

Then process the extracted text as you would any markdown seed.

Procedure

1. Process Findings, Not Seeds

After the ingest pass is complete, enrichment reads from the findings JSON, not the seeds themselves. For each workspace-artifacts/wiki/findings/<citation_id>.json:

Read the JSON and identify which wiki pages it should populate or create
Copy verbatim quotes, equations, and locators into the matching wiki pages
Do NOT re-read the original seed — the seed-reader already extracted it
If a finding is marked completeness: "partial", flag the wiki page's Source Notes with the partial_reason so humans know the extraction is incomplete
If context is nearing limits, persist progress and mark the findings index entry's used_in_wiki: true only for pages you have fully populated
Note: findings are derived directly from SME-curated seeds and are schema-validated — they are the trustworthy source of truth for enrichment

2. Enrich Wiki Pages During Each File Pass

As each seed is processed, update existing wiki pages:

Replace placeholder text with extracted content
Fill in the Definition section with precise, 2-3 sentence definitions
Fill in Formalism with LaTeX math where applicable
Add Key Claims with citation ID references [src-xxx]
Map Relationships (prerequisite_for, depends_on, contradicts, extends)
List Open Questions — what the seed doesn't answer
Add Source Notes with verbatim extractions and page numbers

3. Create New Wiki Pages As Needed

If the current seed file contains concepts not yet in the wiki:

Create new .md files in workspace-artifacts/wiki/pages/
Follow the wiki page schema (see below)
Register citations in workspace-artifacts/wiki/citations/index.yaml

4. Cross-Link Concepts

After processing each seed (and again at the end):

Update related fields in frontmatter
Update Relationships sections with actual connections
Ensure no orphan pages (every page links to at least one other)

5. Validate

Validation runs automatically after the ingest-orchestrator subagent stops (via the subagent_stop_dispatch hook chain). No manual command needed.

Wiki Page Schema

---
id: unique-slug
type: concept | relationship | equation | source | open-question
created: ISO-8601
sources: [citation-id-1, citation-id-2]
related: [concept-id-1, concept-id-2]
status: raw | reviewed | validated
---

Required sections: Definition, Formalism, Key Claims, Relationships, Open Questions, Source Notes.

Citation Format

When adding citations to workspace-artifacts/wiki/citations/index.yaml:

citations:
  src-smith2024-psf:
    human: "Smith et al. (2024), section 3.2"
    source:
      type: seed
      path: /seeds/smith2024_psf_modeling.pdf
      page: 7
      section: "3.2"
    metadata:
      authors: ["Smith", "Jones"]
      title: "PSF Modeling for Space Imaging"
      year: 2024
      venue: "SPIE"
    status: raw

Output

Enriched wiki pages with real extracted content (full text, not summaries)
Updated citation index with page/section/line references for every claim
Updated wiki index and log (rebuild with CLI if needed)
Immediately hand off to prompts/stage-1a2-orchestration.prompt.md and the provisioned .github/agents/stage-1a2-orchestrator.agent.md so Stage 1B/1C loop execution is managed from a single orchestration prompt and real custom agents

Guiding Principles

Document everything so it is auditable by humans and LLMs alike.
Data over folklore — a reference citation is not enough. Include quoted text or locators.
Accessible to everyone — write wiki pages in clear language that a non-expert can follow.
Domain agnostic — do not assume the user's field. This process works for any domain.
Knowledge should be shared — structure content so it can be reused, extended, and challenged.

# Stage 1A: Breadth Research — Prompt Instructions ## Intent **Build an LLM-accessible knowledge base to make an LLM a domain and problem-space expert before any task is posited.** Stage 1A is where raw materials become structured knowledge. You transform seed documents into wiki pages that contain the *actual content* — equations, claims, data, methods — not summaries. **Data over folklore.** A reference citation is not enough. There must be quoted text, page numbers, section numbers, or line numbers. If you cannot point to where a claim comes from, it does not belong in the wiki. ## Your Role Research Crawler agent. You ingest seed documents and build the shared wiki surface. ## Context The human has provided seed documents in `workspace-artifacts/seeds/` and written `PROBLEM_STATEMENT.md`. This prompt is the operator entry point for Stage 1A. Stage 1A now runs in two passes: 1. **Ingest pass** (full-fidelity extraction into findings JSON). 2. **Enrichment pass** (wiki pages built from findings JSON, not from seeds). > Ingest runs automatically on prompt invocation (`meta-compiler ingest --scope all` fires via the `user_prompt_submit_dispatch` hook). Conduct the seed fan-out via the `ingest-orchestrator` subagent. After it stops, `meta-compiler ingest-validate`, `meta-compiler research-breadth`, and `meta-compiler validate-stage --stage 1a` run automatically via the `subagent_stop_dispatch` hook. The `ingest-orchestrator` fans out `seed-reader` subagents, one per seed, and writes one findings JSON per seed under `workspace-artifacts/wiki/findings/<citation_id>.json` plus `findings/index.yaml`. Your enrichment job then reads the findings JSON (not the seeds) to fill wiki pages. The findings JSON already contains verbatim quotes, equations, and locators — treat it as the source of truth for Stage 1A enrichment. ## Delegation Rule - Use the `ingest-orchestrator` agent for reading seeds. It delegates to `seed-reader` subagents, which read each document end-to-end. **DO NOT** use the `explore` subagent to read seeds; Explore hallucinates on long documents. - Use `explore` only for fast scans of wiki pages, reports, or findings JSON that already exist on disk. - Use `research` only when a finding reveals a meaningful missing concept that requires deeper external investigation. - Keep long discovery output in files and wiki artifacts rather than bloating the active orchestration context. ## Critical Rule: Full Paper Text, Not Summaries **Wrong:** "Paper X discusses sensor noise modeling." **Right:** "Paper X establishes that read noise follows Poisson-Gaussian mixture (Eq. 12), parameterized by gain k and offset sigma_read. They validate against EMVA1288 standard (Table 3, p.15)." The difference matters. A summary requires re-reading the paper. A document is reusable by downstream agents. **Enforcement:** Every wiki page must include direct quotes or specific references (page number, section number, equation number, table number, line number) from the source material. Pages that contain only paraphrased summaries without specific locators will be flagged during validation. **Non-plaintext seeds:** If a seed is a PDF, DOCX, XLSX, or PPTX, extract its text first: ```bash python scripts/read_document.py workspace-artifacts/seeds/paper.pdf --output /tmp/paper_text.md ``` Then process the extracted text as you would any markdown seed. ## Procedure ### 1. Process Findings, Not Seeds After the ingest pass is complete, enrichment reads from the findings JSON, not the seeds themselves. For each `workspace-artifacts/wiki/findings/<citation_id>.json`: - Read the JSON and identify which wiki pages it should populate or create - Copy verbatim quotes, equations, and locators into the matching wiki pages - Do NOT re-read the original seed — the seed-reader already extracted it - If a finding is marked `completeness: "partial"`, flag the wiki page's Source Notes with the `partial_reason` so humans know the extraction is incomplete - If context is nearing limits, persist progress and mark the findings index entry's `used_in_wiki: true` only for pages you have fully populated - Note: findings are derived directly from SME-curated seeds and are schema-validated — they are the trustworthy source of truth for enrichment ### 2. Enrich Wiki Pages During Each File Pass As each seed is processed, update existing wiki pages: - Replace placeholder text with extracted content - Fill in the Definition section with precise, 2-3 sentence definitions - Fill in Formalism with LaTeX math where applicable - Add Key Claims with citation ID references `[src-xxx]` - Map Relationships (prerequisite_for, depends_on, contradicts, extends) - List Open Questions — what the seed doesn't answer - Add Source Notes with verbatim extractions and page numbers ### 3. Create New Wiki Pages As Needed If the current seed file contains concepts not yet in the wiki: - Create new `.md` files in `workspace-artifacts/wiki/pages/` - Follow the wiki page schema (see below) - Register citations in `workspace-artifacts/wiki/citations/index.yaml` ### 4. Cross-Link Concepts After processing each seed (and again at the end): - Update `related` fields in frontmatter - Update Relationships sections with actual connections - Ensure no orphan pages (every page links to at least one other) ### 5. Validate Validation runs automatically after the `ingest-orchestrator` subagent stops (via the `subagent_stop_dispatch` hook chain). No manual command needed. ## Wiki Page Schema ```yaml --- id: unique-slug type: concept | relationship | equation | source | open-question created: ISO-8601 sources: [citation-id-1, citation-id-2] related: [concept-id-1, concept-id-2] status: raw | reviewed | validated --- ``` Required sections: Definition, Formalism, Key Claims, Relationships, Open Questions, Source Notes. ## Citation Format When adding citations to `workspace-artifacts/wiki/citations/index.yaml`: ```yaml citations: src-smith2024-psf: human: "Smith et al. (2024), section 3.2" source: type: seed path: /seeds/smith2024_psf_modeling.pdf page: 7 section: "3.2" metadata: authors: ["Smith", "Jones"] title: "PSF Modeling for Space Imaging" year: 2024 venue: "SPIE" status: raw ``` ## Output - Enriched wiki pages with real extracted content (full text, not summaries) - Updated citation index with page/section/line references for every claim - Updated wiki index and log (rebuild with CLI if needed) - Immediately hand off to `prompts/stage-1a2-orchestration.prompt.md` and the provisioned `.github/agents/stage-1a2-orchestrator.agent.md` so Stage 1B/1C loop execution is managed from a single orchestration prompt and real custom agents ## Guiding Principles - **Document everything** so it is auditable by humans and LLMs alike. - **Data over folklore** — a reference citation is not enough. Include quoted text or locators. - **Accessible to everyone** — write wiki pages in clear language that a non-expert can follow. - **Domain agnostic** — do not assume the user's field. This process works for any domain. - **Knowledge should be shared** — structure content so it can be reused, extended, and challenged.

{ "prompt": "# Stage 1A: Breadth Research — Prompt Instructions\n\n## Intent\n\n**Build an LLM-accessible knowledge base to make an LLM a domain and problem-space\nexpert before any task is posited.** Stage 1A is where raw materials become\nstructured knowledge. You transform seed documents into wiki pages that contain\nthe *actual content* — equations, claims, data, methods — not summaries.\n\n**Data over folklore.** A reference citation is not enough. There must be quoted\ntext, page numbers, section numbers, or line numbers. If you cannot point to where\na claim comes from, it does not belong in the wiki.\n\n## Your Role\nResearch Crawler agent. You ingest seed documents and build the shared wiki surface.\n\n## Context\nThe human has provided seed documents in `workspace-artifacts/seeds/` and written\n`PROBLEM_STATEMENT.md`. This prompt is the operator entry point for Stage 1A.\n\nStage 1A now runs in two passes:\n\n1. **Ingest pass** (full-fidelity extraction into findings JSON).\n2. **Enrichment pass** (wiki pages built from findings JSON, not from seeds).\n\n> Ingest runs automatically on prompt invocation (`meta-compiler ingest --scope all` fires via the `user_prompt_submit_dispatch` hook). Conduct the seed fan-out via the `ingest-orchestrator` subagent. After it stops, `meta-compiler ingest-validate`, `meta-compiler research-breadth`, and `meta-compiler validate-stage --stage 1a` run automatically via the `subagent_stop_dispatch` hook.\n\nThe `ingest-orchestrator` fans out `seed-reader` subagents, one per seed, and\nwrites one findings JSON per seed under\n`workspace-artifacts/wiki/findings/<citation_id>.json` plus\n`findings/index.yaml`.\n\nYour enrichment job then reads the findings JSON (not the seeds) to fill wiki\npages. The findings JSON already contains verbatim quotes, equations, and\nlocators — treat it as the source of truth for Stage 1A enrichment.\n\n## Delegation Rule\n- Use the `ingest-orchestrator` agent for reading seeds. It delegates to\n `seed-reader` subagents, which read each document end-to-end. **DO NOT** use\n the `explore` subagent to read seeds; Explore hallucinates on long documents.\n- Use `explore` only for fast scans of wiki pages, reports, or findings JSON\n that already exist on disk.\n- Use `research` only when a finding reveals a meaningful missing concept that\n requires deeper external investigation.\n- Keep long discovery output in files and wiki artifacts rather than bloating\n the active orchestration context.\n\n## Critical Rule: Full Paper Text, Not Summaries\n\n**Wrong:** \"Paper X discusses sensor noise modeling.\"\n**Right:** \"Paper X establishes that read noise follows Poisson-Gaussian mixture\n(Eq. 12), parameterized by gain k and offset sigma_read. They validate against\nEMVA1288 standard (Table 3, p.15).\"\n\nThe difference matters. A summary requires re-reading the paper. A document is\nreusable by downstream agents.\n\n**Enforcement:** Every wiki page must include direct quotes or specific references\n(page number, section number, equation number, table number, line number) from the\nsource material. Pages that contain only paraphrased summaries without specific\nlocators will be flagged during validation.\n\n**Non-plaintext seeds:** If a seed is a PDF, DOCX, XLSX, or PPTX, extract its\ntext first:\n```bash\npython scripts/read_document.py workspace-artifacts/seeds/paper.pdf --output /tmp/paper_text.md\n```\nThen process the extracted text as you would any markdown seed.\n\n## Procedure\n\n### 1. Process Findings, Not Seeds\nAfter the ingest pass is complete, enrichment reads from the findings JSON, not\nthe seeds themselves. For each `workspace-artifacts/wiki/findings/<citation_id>.json`:\n- Read the JSON and identify which wiki pages it should populate or create\n- Copy verbatim quotes, equations, and locators into the matching wiki pages\n- Do NOT re-read the original seed — the seed-reader already extracted it\n- If a finding is marked `completeness: \"partial\"`, flag the wiki page's\n Source Notes with the `partial_reason` so humans know the extraction is incomplete\n- If context is nearing limits, persist progress and mark the findings index\n entry's `used_in_wiki: true` only for pages you have fully populated\n- Note: findings are derived directly from SME-curated seeds and are\n schema-validated — they are the trustworthy source of truth for enrichment\n\n### 2. Enrich Wiki Pages During Each File Pass\nAs each seed is processed, update existing wiki pages:\n- Replace placeholder text with extracted content\n- Fill in the Definition section with precise, 2-3 sentence definitions\n- Fill in Formalism with LaTeX math where applicable\n- Add Key Claims with citation ID references `[src-xxx]`\n- Map Relationships (prerequisite_for, depends_on, contradicts, extends)\n- List Open Questions — what the seed doesn't answer\n- Add Source Notes with verbatim extractions and page numbers\n\n### 3. Create New Wiki Pages As Needed\nIf the current seed file contains concepts not yet in the wiki:\n- Create new `.md` files in `workspace-artifacts/wiki/pages/`\n- Follow the wiki page schema (see below)\n- Register citations in `workspace-artifacts/wiki/citations/index.yaml`\n\n### 4. Cross-Link Concepts\nAfter processing each seed (and again at the end):\n- Update `related` fields in frontmatter\n- Update Relationships sections with actual connections\n- Ensure no orphan pages (every page links to at least one other)\n\n### 5. Validate\nValidation runs automatically after the `ingest-orchestrator` subagent stops\n(via the `subagent_stop_dispatch` hook chain). No manual command needed.\n\n## Wiki Page Schema\n\n```yaml\n---\nid: unique-slug\ntype: concept | relationship | equation | source | open-question\ncreated: ISO-8601\nsources: [citation-id-1, citation-id-2]\nrelated: [concept-id-1, concept-id-2]\nstatus: raw | reviewed | validated\n---\n```\n\nRequired sections: Definition, Formalism, Key Claims, Relationships,\nOpen Questions, Source Notes.\n\n## Citation Format\n\nWhen adding citations to `workspace-artifacts/wiki/citations/index.yaml`:\n\n```yaml\ncitations:\n src-smith2024-psf:\n human: \"Smith et al. (2024), section 3.2\"\n source:\n type: seed\n path: /seeds/smith2024_psf_modeling.pdf\n page: 7\n section: \"3.2\"\n metadata:\n authors: [\"Smith\", \"Jones\"]\n title: \"PSF Modeling for Space Imaging\"\n year: 2024\n venue: \"SPIE\"\n status: raw\n```\n\n## Output\n- Enriched wiki pages with real extracted content (full text, not summaries)\n- Updated citation index with page/section/line references for every claim\n- Updated wiki index and log (rebuild with CLI if needed)\n- Immediately hand off to `prompts/stage-1a2-orchestration.prompt.md` and the provisioned `.github/agents/stage-1a2-orchestrator.agent.md` so Stage 1B/1C loop execution is managed from a single orchestration prompt and real custom agents\n\n## Guiding Principles\n- **Document everything** so it is auditable by humans and LLMs alike.\n- **Data over folklore** — a reference citation is not enough. Include quoted text or locators.\n- **Accessible to everyone** — write wiki pages in clear language that a non-expert can follow.\n- **Domain agnostic** — do not assume the user's field. This process works for any domain.\n- **Knowledge should be shared** — structure content so it can be reused, extended, and challenged.", "variables": [], "notes": "Stage 1A: Breadth Research — Prompt Instructions", "metadata": { "source": { "provider": "github-prompts", "repository": "https://github.com/cstokes0091/META-COMPILER", "path": ".github/prompts/stage-1a-breadth.prompt.md", "ref": "dd3719f133ccf11bf033adfed1ad2f002eedb7bc", "url": "https://github.com/cstokes0091/META-COMPILER/blob/dd3719f133ccf11bf033adfed1ad2f002eedb7bc/.github/prompts/stage-1a-breadth.prompt.md", "key": "cstokes0091/META-COMPILER/.github/prompts/stage-1a-breadth.prompt.md" } }, "model_hints": [ "GPT-5.2 (copilot)" ] }

stage-1a-breadth.prompt

About