Painpoint To MVP#

Modes#

1. Discovery mode: find and rank product opportunities from public user discussion evidence.
2. PRD mode: turn one persisted opportunity into a lightweight PRD for a coding agent.

If the user asks for both, discovery comes first.

Hard Rules#

• references/contracts.md is the only runtime contract reference.
• Only inspect scripts when a documented helper command fails unexpectedly or when explicitly modifying this skill.
• Read references/contracts.md once per run, not repeatedly.

Preferred Commands#

Use these commands for normal operation:

• npm run bootstrap:tidb
• npm run tidb -- save-evidence-batch
• npm run tidb -- save-opportunity-batch
• npm run tidb -- get-opportunity
• npm run tidb -- save-prd
• npm run tidb -- get-prd
• npm run tidb -- list-runs

Compatibility-only commands:

• npm run tidb -- save-evidence
• npm run tidb -- save-opportunity

Use the single-row save commands only for manual recovery or compatibility flows.

Normal Execution Defaults#

• Check required environment once per run.
• Collect evidence first, then persist in batch.
• Prefer JSON via stdin over temp files.
• Use --input-file only when the payload is too large or awkward for stdin.

Discovery Mode#

1. Read references/contracts.md once.
2. Check required environment.
3. Collect evidence with search/scraping tools or skills (e.g. bundled websearch, meirkad/bright-data, buksan1950/reddit-readonly).
4. Persist all evidence with save-evidence-batch.
5. Cluster and rank opportunities using the contract rubric.
6. Persist ranked opportunities with save-opportunity-batch.
7. Return a ranked summary with run_id, opportunity_id, score, confidence, and caveats.

PRD Mode#

1. Resolve the target opportunity by opportunity_id, or by run_id plus rank when explicitly provided.
2. Read the persisted opportunity with get-opportunity.
3. Generate the PRD from the persisted opportunity and its evidence, not chat memory.
4. Persist the PRD with save-prd.
5. Return the human-readable PRD and include prd_id.

Non-Drift Rules#

• Never invent SQL. Use the documented helper commands.
• Never generate a PRD from an unpersisted opportunity summary.
• Never use conversational numbering like "opportunity #2" as the only identifier once results are persisted.
• Never claim product-market fit.
• When evidence is weak, say insufficient evidence.
• Preserve contradictions when sources disagree.

Quality Bar#

• Prefer repeated, concrete complaints over clever one-offs.
• Distinguish category pain from vendor-specific complaints.
• Keep evidence traceable through source_url, pain_cluster_id, and persisted identifiers.
• Optimize PRDs for a coding agent with explicit scope, non-goals, constraints, and acceptance criteria.