context-authoring

context-authoring

When to use

Use this skill when:

• A new project has been scaffolded and the agents/settings/contexts/ files are still template stubs from event4u/agent-config.
• The user asks "help me fill in the auth model context", "set up tenant boundaries", or similar knowledge-layer work.
• A reviewer skill (authz-review, data-flow-mapper, migration-safety, multi-tenant-boundary-review, secrets-and-config-review) reports "I cannot proceed — agents/settings/contexts/<file>.md is still a template".
• After a significant architecture change that invalidates one of the five context files.

Do NOT use when:

• Writing a regular roadmap or feature doc — use agent-docs-writing.
• Creating a generic context document that does not map to one of the five knowledge-layer templates — use context-create.
• Filling in the engineering-memory YAML files (domain-invariants.yml, etc.) — use /memory-add.

The five files

The templates ship in .agent-src.uncondensed/templates/contexts/ and are copied into agents/settings/contexts/ by the installer.

Procedure: context-authoring

Step 0: Inspect

1. List agents/settings/contexts/ — which of the five files exist? Which still contain the <!-- Template shipped by event4u/agent-config. --> HTML comment?

2. Ask the user which file to work on. Use numbered options:

   > 1. auth-model.md — roles, permissions, impersonation
   > 2. tenant-boundaries.md — tenancy type and scope propagation
   > 3. data-sensitivity.md — field classification and masking
   > 4. deployment-order.md — migrations, flags, rollback
   > 5. observability.md — errors, logs, metrics, alerts
   

3. If multiple files are stubs, default to the order above — auth-model is the prerequisite for tenant-boundaries; both feed data-sensitivity.

Step 1: Harvest evidence before asking

For the chosen file, pull what the codebase already reveals before asking the user. Record the file:line citations — they become the authoritative source when the user is unsure.

Start the walkthrough by showing the harvested evidence — the user only has to confirm or correct, not invent from scratch.

Step 2: Walk the template section by section

1. Open the template and treat every HTML comment as a question for the user. Do NOT fabricate answers to skip a section.

2. Present each section as:

   **Section:** <heading>
   **Template asks:** <what the comment says>
   **Evidence I found:** <file:line references or "none">
   **Proposed content:** <draft or "I need your input">
   
   > 1. Accept the draft
   > 2. Edit — tell me what's wrong
   > 3. I don't know — mark as "TBD" with a follow-up task
   

3. If the user picks "TBD", insert an HTML comment <!-- TBD: <question> --> at that spot — never a fabricated value. Reviewer skills key on the comment to warn about incomplete sections.

Step 3: Preserve the file contract

1. Remove the top-of-file <!-- Template shipped by ... --> comment only after at least one section has been authored — an untouched file must stay recognisable as a stub.
2. Keep every heading the template ships with. Reviewer skills grep for exact section names (## Known exceptions, ## Known alerts, etc.).
3. Do NOT add new top-level sections. The template surface is the contract — extend existing sections, file a proposal via learning-to-rule-or-skill to expand the template upstream.

Step 4: Validate

1. Run ./scripts-run src/scripts/check_portability — project-specific content is expected here, but the check catches accidental copy of other projects' identifiers.
2. Run ./scripts-run src/scripts/check_references — cross-file links between the five contexts must resolve.
3. Confirm with the user: "is this accurate enough that a reviewer should treat it as the source of truth?" — anything less and a <!-- TBD --> marker stays.

Output format

1. agents/settings/contexts/<file>.md updated with project-specific content; every section either authored or explicitly marked <!-- TBD: ... -->.
2. A short summary comment back to the user: which sections are complete, which are TBD, and which downstream reviewer skills are now unblocked.
3. Optional: a proposal stub in agents/learnings/ for any template gap the user hit (missing section, ambiguous field) — feeds the curated self-improvement pipeline via learning-to-rule-or-skill.

Gotcha

• The model tends to fabricate plausible roles, fields, or alerts when harvesting comes up empty. Do NOT. An <!-- TBD: ... --> marker is always better than a made-up entry — reviewer skills trust this file.
• The model tends to collapse the template once it starts editing, losing the HTML comments that explain why a section exists. Preserve them until the section is authored — they are the authoring prompt.
• data-sensitivity.md is the highest-leverage file and also the one most likely to be skipped as "boring". Prioritise it after auth-model — missing entries here become production leaks, not review nits.
• Do not treat this skill as a form-filler. If a project is single-tenant, tenant-boundaries.md SHOULD be deleted, not stubbed — the checklist explicitly says so at the top of the file.

Path conventions when a context cites another context

If the context file declares its own load_context: (chain reasoning), use logical names rooted at the source — contexts/<area>/<file>.md for package material or agents/settings/contexts/<file>.md for project-local material. The .agent-src.uncondensed/ prefix is rejected by the schema regex and by scripts/lint_load_context.ts. Body links to docs/guidelines/... use the verbatim ../../docs/... relative form. Canonical reference: rule-writing § 3b and docs/contracts/load-context-schema.md.

Frugality Standards

Apply the Frugality Charter to every context file you author.

Examples in this artifact:

• Per the charter's index nature, context files are reviewer fuel — they hold mechanics, not Iron-Law obligations.
• Per the cite-don't-restate principle, when a section mirrors a rule, link the rule and stop.
• Per the act-skip-narration rule, lookup tables open the section; explanatory prose follows only if the table is ambiguous.

Pre-save self-check:

1. Does the context file restate Iron-Law text instead of linking the rule?
2. Does any section open with "This document explains…" instead of the lookup material?
3. Are placeholders (<add me>, TBD) shipped instead of actual content?
4. Are the cited rules linked with stable anchors (verified to exist)?

Do NOT

• Do NOT copy content between projects. Every context file is local to its repo. Reuse of another project's roles, tenants, or alerts is a portability violation and a security risk.
• Do NOT commit TBD-heavy files without flagging them in the PR description. Reviewer skills will downgrade confidence, but a human reviewer should know the contexts are partial.
• Do NOT rename or restructure the template sections. Reviewer skills grep for exact headings.