context-probe-analysis-workflow

Context Probe Analysis Workflow

Run context-probe after curated inputs already exist.

Respond in English while this skill is active unless the user explicitly asks for another language.

If the repo does not have maintained YAML inputs yet, switch to context-probe-inputs-workflow first.

Read references/command-recipes.md when you need exact command shapes. Read references/interpret-results.md immediately after any run. Read references/investigate-correctness.md only when the result looks suspicious or too weak to trust. Read references/final-summary.md before closing.

Workflow

1. Classify the request

Place the user in one of these paths before running anything:

• fresh domain_design measurement
• fresh architecture_design measurement
• readable report or gate decision from existing inputs
• unknown or low-confidence review
• suspicious result investigation
• explicit review resolution replay

2. Start with the smallest useful path

Run the narrowest command that answers the request:

• first-time or broad measurement: start with score.compute
• human-readable output: add report.generate
• threshold decision only: use gate.evaluate
• human review targets: use review.list_unknowns
• extraction, trace, and history commands: use only to inspect why a result looks weak or wrong
• review.resolve: use only when review items and explicit human resolutions already exist

3. Interpret before escalating

Read outputs in this order:

1. status
2. summary of result
3. confidence
4. unknowns
5. diagnostics
6. provenance

Separate the result into:

• metrics already directional enough to discuss
• metrics still proxy-heavy because inputs or observations are missing
• mismatches that suggest wrong inputs or wrong expectations

4. Investigate trust and correctness

If results look suspicious, incomplete, or lower-confidence than expected, inspect likely causes before changing curated YAML:

• use review.list_unknowns to see unresolved gaps
• use doc.extract_* to inspect extracted terms, rules, or invariants
• use trace.link_terms or trace.link_model_to_code to inspect linking evidence
• use history.* to inspect locality and co-change signals
• use review.resolve only to apply explicit human review decisions, then rerun the affected extraction or score path

5. Close with a disposition

End with a concise summary that states:

• what was measured
• what looks high confidence
• what remains unknown
• recommended next actions
• the next command only if it materially advances the analysis

Do not turn every run into an issue proposal. Finish with the measured state first.

Guardrails

• Do not use this skill to invent or curate YAML inputs from scratch.
• Do not start with doc.extract_* for a first measurement unless direct evidence inspection is the task.
• Do not confuse --model and --constraints.
• Do not treat status: ok as proof that the result is complete or high confidence.
• Do not ignore unknowns, diagnostics, or provenance.
• Do not edit curated YAML only to suppress unknowns. Inspect the evidence path first.