AI Commenting

AI Commenting Skill

Native Subagent Protocol (Codex)

Codex supports native subagents. Delegate with spawn_agent, coordinate with send_input, collect via wait_agent, and clean up with close_agent.

Execution preference:

1. Use native subagents first for independent workstreams (parallel when possible).
2. Merge results in main thread and run final verification.
3. Fallback only when delegation is blocked: use the [ANALYST]/[ARCHITECT]/[EXECUTOR]/[REVIEWER] structure in a single response.

Minimal orchestration pattern:

spawn_agent -> send_input (optional) -> wait_agent -> close_agent

Create and maintain AI-native annotations so models can understand project intent quickly and safely, not just code syntax.

Why This Skill Exists

Traditional comments are optimized for human reading, but LLM coding agents need dense, structured context to avoid unsafe edits and repeated discovery work.

This skill defines a compact annotation protocol that turns files into a machine-readable context layer for:

• Faster onboarding into unfamiliar modules
• Risk-aware edit strategy
• Better test planning
• Lower regression rate in high-coupling code

Research-Backed Principles

Use these principles when writing AI annotations:

• Explain why and constraints, not obvious what.
• Keep instructions explicit: goal, constraints, and acceptance checks.
• Standardize structure for easy parsing and retrieval.
• Keep annotations close to code and update them with code changes.
• Prefer short, high-signal metadata over long prose.

Annotation Format

Canonical format (single line):

/*@ai:key=value|key=value|key=value*/

Rules:

• Use ASCII keys and values.
• Separate fields with |.
• No spaces around =.
• Use concise tokens (AuthService, integration, p95<200ms).
• Keep one annotation to one scope (file or block).

Field Schema

Core fields (recommended at file level):

• risk=1-5: change risk (5 is highest)
• core=<domain>: core responsibility (UserCRUD, BillingLedger)
• deps=<A,B,C>: critical dependencies
• intent=<why>: non-obvious business intent
• test=<unit|integration|e2e|contract>: minimum test gate

Extended fields (use when needed):

• chain=<A->B->C>: business or data flow chain
• async=<low|medium|complex>: async/concurrency complexity
• api=<internal|external>: API boundary type
• auth=<none|required|strict>: authorization level
• invariant=<must_hold>: critical invariant
• sidefx=<db,cache,queue,event>: side effects
• perf=<budget>: performance constraint (p95<200ms)
• security=<pii|payment|secret|none>: security sensitivity
• owner=<team_or_module>: ownership hint
• rollback=<strategy>: safe fallback strategy
• ticket=<id>: change traceability

Optional maintenance fields:

• lines=<N|500+>: file size hint
• updated=<YYYY-MM-DD>: annotation freshness

Risk Rubric (Required if risk present)

• risk=1: isolated logic, low coupling, fast rollback
• risk=2: local impact, clear tests, low side effects
• risk=3: moderate coupling or external contract dependency
• risk=4: cross-module critical path, data/auth impact
• risk=5: security/payment/core-state changes, high blast radius

Escalate risk by +1 when any condition applies:

• touches auth/session/payment/PII/secrets
• changes distributed state consistency
• modifies external API contract
• lacks reproducible integration tests

LLM Context Checklist (Project Coding)

For high-value modules, ensure annotations expose these context dimensions:

• Business goal and intent (intent, core)
• Hard constraints and invariants (invariant, perf, security)
• Dependency and flow graph (deps, chain, sidefx)
• Interface contract and auth boundary (api, auth)
• Validation strategy (test)
• Operational safety (risk, rollback, owner)

If the model can read these fields, it can plan edits with less guesswork.

Placement Strategy

Use a two-layer strategy:

1. File header annotation (always for critical modules)

• Place near imports or module declaration.
• Include at least: risk, core, deps, intent, test.

2. Critical block annotation (selective)

• Place above risky functions/flows.
• Include only fields needed for that block (invariant, auth, sidefx, perf).

Do not annotate every function. Keep density high-signal.

Your Draft, Upgraded

Example from your draft (fully valid):

/*@ai:risk=4|deps=UserModel,AuthService,Database|lines=500+|core=UserCRUD*/
/*@ai:chain=Auth->User->Permission->Audit|async=complex|test=integration*/
class UserManager {
  // ...
}

Recommended enriched version:

/*@ai:risk=5|core=UserCRUD|intent=protect_user_consistency|deps=UserModel,AuthService,AuditService*/
/*@ai:chain=Auth->User->Permission->Audit|auth=strict|sidefx=db,event|test=integration|rollback=feature_flag*/
class UserManager {
  /*@ai:risk=4|invariant=user_id_unique|security=pii|perf=p95<200ms*/
  async deleteUser(userId: string) {
    // ...
  }
}

Execution Workflow

When asked to annotate a module:

1. Discover scope

• Identify files and high-risk boundaries.
• Gather dependency and contract points.

2. Assign risk and context fields

• Start with file-level core fields.
• Add block-level fields only for risky paths.

3. Validate syntax and quality

• Ensure parseable /*@ai:...*/ format.
• Remove duplicate or contradictory tags.

4. Tie to tests and safety

• Ensure test and rollback are present for risk>=4.

5. Output concise report

• Files touched
• Risk distribution
• Missing test/safety gaps

Quality Gates

Minimum acceptance:

• All risk>=4 files have file-level @ai annotations.
• Each risk>=4 block includes test or inherits explicit file-level test.
• security!=none implies auth!=none and a rollback strategy.
• No stale tags older than 90 days without verification (updated= recommended).

Density guidance:

• Prefer 1 annotation per 60-120 LOC.
• Avoid repetitive tags on low-risk utility code.

Anti-Patterns

Avoid:

• Restating code mechanics as metadata.
• Conflicting tags in one scope (auth=none and security=pii).
• Unbounded text blobs in intent.
• Tag sprawl without test or rollback fields.

Parsing Helpers

Use these regex patterns when extracting annotations:

const aiTagPattern = /\/\*@ai:([^*]+)\*\//g;
const fieldPattern = /(\w+)=([^|*]+)/g;

Normalize output to key-value JSON for downstream retrieval and risk maps.

Deliverable Template

When this skill completes, return:

• Result summary: what annotation protocol was applied
• Files changed: absolute file paths
• Validation evidence: parser/lint/check commands and outcomes
• Risks / next actions: unresolved safety/test gaps

References

• OpenAI prompt engineering best practices: clear instruction, context, constraints, and evaluation criteria
• Anthropic prompt engineering: explicit structure and measurable output constraints
• GitHub Copilot custom instructions: repository-specific standards and conventions improve coding quality
• Google style guides: comments should prioritize intent and maintainability