Write Financial Research Others Can Recheck, in a Project Shared by Several AI Tools

## Decide Which File Is the Main Copy and Which Are Tool Copies Map repository responsibilities before editing. Distinguish canonical workflow sources, generated agent-specific packages, generated prompt wrappers, shared financial tools, research reports, installer scripts, and runtime-specific instruction files. Keep one declared workflow source authoritative across supported agent environments. Treat generated packages and prompt wrappers as compatibility outputs, not independent sources of truth. Keep each runtime's behavior file separate so platform-specific instructions do not silently alter another environment. ## Rebuild Each AI Tool's Copy After Changing the Main Workflow After changing a canonical workflow, run the repository's synchronization process to regenerate every required agent package. Regenerate prompt-style wrappers only when that compatibility surface is supported or requested; prefer full workflow packages when both entry points exist. Do not manually edit a generated package without also updating its canonical source. Before completion, use a check-only synchronization mode when available to prove that generated artifacts are current without rewriting them. Review the generated diff so synchronization does not conceal unrelated changes. ## Keep Files Meant for Only One AI Tool Clearly Separate Allow a hand-written, platform-only package only when it is clearly marked and no same-named canonical workflow exists. If the workflow is intentionally adopted by multiple environments, move its shared behavior into the canonical source and regenerate the platform adapters. Provide installation paths for each supported operating system and keep documented tool locations consistent with the expected checkout layout. Test the compatibility path relevant to the change rather than assuming one environment's successful installation proves another's. ## Fix the Research Date Before Starting and Show It in the Report Before researching: 1. Read the current date from the execution environment. 2. Use that date as the baseline for terms such as latest, current, and most recent. 3. State the data cutoff date in the report header. 4. Distinguish information available by the cutoff from events learned afterward. Never infer the current date from model training data. When sources update at different times, record the observation time or reporting period needed to explain the difference. ## Check Each Financial Fact Against Two Independent Sources Use at least two independent sources when a claim requires verification. Prefer primary filings, regulatory records, issuer disclosures, and direct market data for core facts; use aggregators and news as corroboration or discovery aids rather than silent substitutes. Record the source, publication or retrieval date, currency, units, reporting period, share basis, and adjustment basis needed for comparison. Reconcile disagreements explicitly. Do not average conflicting values or choose the more convenient number without explaining why one source is more applicable. ## Do Financial Math With Calculation Tools So Others Can Repeat It Use deterministic calculation tools for market capitalization, valuation multiples, cross-source checks, and scenario analysis. Preserve the input values, formulas, units, rounding policy, and output so another reviewer can reproduce the result. Normalize currencies, per-share versus aggregate values, diluted versus basic share counts, and fiscal versus calendar periods before comparison. Keep assumptions separate from observed data, and run sensitivity cases when the conclusion depends materially on an uncertain input. ## Mark Weak Spots and Avoid Telling People to Buy or Sell Separate sourced facts, calculated results, assumptions, and analyst inferences. Clearly label low-confidence conclusions, incomplete data, stale inputs, conflicting sources, and unresolved gaps. Do not present an unsupported estimate as a verified current value. Frame the work as learning and research rather than investment advice. Avoid unsupported directives to buy, sell, or hold. State the limitations that could change the conclusion and identify the evidence required to raise confidence. ## Leave Existing Reports Alone and Change Only What Was Asked Preserve existing report files unless the task specifically requests a report change. A tooling, workflow, or compatibility update must not rewrite unrelated research outputs. Limit changes to the requested workflow, tool, script, documentation, or report. Keep generated artifacts in their defined locations and exclude temporary results from the change set. Inspect the focused diff before completion and separate canonical-source edits from their generated consequences. ## Run the Checks and Review the Report Before Calling Research Done Run the relevant syntax, generation, and compatibility checks for every changed artifact. Use non-writing freshness checks when the task only needs to confirm synchronization. Do not claim cross-platform compatibility from a check that exercised only one platform. Before treating research as publishable, run the report audit and confirm the cutoff date, citations, independent-source requirement, exact calculations, units, uncertainty labels, advice boundary, and scope of changes. Report which checks passed, which could not run, and any source or calculation gap that remains.