Product & Project Delivery
Timebox a Technical Spike and Decide What Evidence Is Enough
Set a bounded learning question, time ceiling and rejection condition. A completed spike plan and real local synthetic search experiment show how evidence changes the implementation judgment and estimate basis.
Timebox a technical spike around one decision, a fixed investigation budget and an explicit evidence threshold. State what would support the approach, what would reject it, and what you will record if the time expires with the question unresolved. The deliverable is a technical judgment supported by recorded evidence, with production readiness and delivery estimates assessed separately.
A useful spike can end with a failed feasibility condition. That result may prevent a team from committing to an unsuitable approach. This guide from LandOffer.ai combines a complete fictional spike plan with a real, narrowly bounded local SQLite experiment. The project, people, planned budget and estimate decisions are invented; the reported query results were actually executed against six synthetic documents. No application, production database or customer account was accessed or changed.

Choose the uncertainty that changes the implementation decision
“Investigate search” can consume a week without producing an answer. “Can the proposed search component distinguish these product codes while supporting the required word queries?” gives the investigation a boundary. The answer affects whether the team can use one search path or needs a different design.
Mike Cohn's explanation of agile spikes describes a spike as timeboxed research that produces knowledge for later work. He recommends using it for excessive uncertainty rather than trying to eliminate the ordinary uncertainty in every task. His examples allow a subsequent decision to investigate further; the original timebox is not silently extended.
Identify who needs the answer and what they would do differently. An engineer may want to understand a library's behavior; the product decision may depend on a stricter condition, such as never showing a different product code as an exact match. Put that condition in the plan before the experiment so reviewers can distinguish a working search from a correct exact-code lookup. Otherwise, a working demo can become the success criterion simply because it is what you managed to build.
Also ask whether existing documentation already settles the question. Use the spike to answer a consequential gap that existing documentation has not resolved for your decision. Reading the documentation may narrow the required experiment to a few counterexamples.
Complete a plan with a stop rule
DeskLeaf is an invented offline help viewer. Engineer Sasha is considering SQLite FTS5 for article search. Product lead Talia needs both ordinary word discovery and exact product-code lookup. In this fictional requirement, AB-12, AB 12 and AB12 are three distinct stored identifiers; returning one for another as an exact match is unacceptable.
The completed plan is intentionally small. Its proposed four-hour ceiling is a teaching assumption, not a recommended universal duration or a claim about how long the local execution took.
| Plan field | Completed fictional agreement |
|---|---|
| Decision | Can a default FTS5 query path satisfy both word discovery and exact-code identification, or must those concerns be separated? |
| Environment and data | Local in-memory SQLite with six authored help documents. Verify FTS5 availability first. No app integration, customer documents, network service or production data |
| Word evidence | Compare cafe, café and the prefix print* against explicit expected document IDs. These cases cover only the chosen fixture |
| Exact-code gate | Querying AB-12 as an exact identifier must return only its document, not the document labeled AB 12. Preserve a counterexample even if other searches work |
| Planned budget | 30 minutes for documentation and environment checks; 90 for the minimal fixture; 60 for one bounded alternative if needed; 60 for evidence and decision notes. Total: four hours |
| Stop conditions | Stop the FTS-only feasibility claim when a decisive code collision is reproduced. If FTS5 is unavailable, record the environment blocker. At the ceiling, stop investigation and report unresolved questions |
| Output and ownership | Sasha supplies the fixture, queries, expected and observed results, limits and recommendation. Talia reviews the requirement interpretation; engineering lead Gabe owns any later implementation estimate |
| Exclusions | No production-readiness claim, performance target, ranking evaluation, update lifecycle, authorization review, migration or release commitment |
The stop rule permits a limited alternative check; it does not authorize redesigning the application inside the spike. If that check exposes another substantial uncertainty, document it for a new decision rather than using the remaining time as permission to keep expanding the work.
The evidence threshold is also asymmetric. One relevant counterexample can reject the claim that this query path preserves exact codes in the fixture. Several successful searches cannot establish that it supports every language, identifier format or deployment environment.
Run the smallest experiment that can expose the mismatch
SQLite's FTS5 documentation explains that queries and document text are tokenized, and that a quoted phrase matches an ordered sequence of tokens. The default Unicode61 tokenizer treats punctuation as separators and removes many Latin-script diacritics. Quotation marks therefore do not turn a full-text phrase into a byte-for-byte identifier comparison.
Executed locally on October 27, 2026, the experiment used Python 3.12.14 and SQLite 3.54.1 with FTS5 available. It created an in-memory FTS5 table using the default unicode61 tokenizer, plus an ordinary table for a separate code-equality check. The six records are entirely synthetic. Three use the codes AB-12, AB 12 and AB12; the others include unrelated codes or no code. Five documents contain cafe or café; two contain a word beginning with print.
The following is the completed evidence record, not a proposed test plan. The runnable fixture and its saved results accompany the local draft.
| Executed check | Actual local observation | What it establishes |
|---|---|---|
FTS query cafe and separate query café |
Both return document IDs 1, 2, 3, 4 and 6 | Accent folding behaves as expected for these specific words and records |
FTS prefix query print* |
Returns IDs 1 and 5 | The chosen prefix query matches the expected fixture records; no latency or ranking claim |
FTS quoted phrase "AB-12" |
Returns IDs 1 and 2, although only ID 1 carries the hyphenated code | The FTS-only path fails the fictional exact-code gate |
| Ordinary equality lookup for each of the three stored codes | AB-12 returns 1; AB 12 returns 2; AB12 returns 3 |
A separate lookup distinguishes these stored uppercase strings in this fixture |
Ordinary equality lookup for ab-12 |
Returns no document | Case policy remains a product/design question; this is not accepted as the desired user experience |
Unquoted FTS expression AB-12 |
Raises the local error no such column: 12 |
Raw identifier text is not automatically a valid FTS query expression; production error handling remains unimplemented |
Nine query cases were executed, with 12 assertions checking the expected local observations and the relevant gate results. The exact-code business condition failed even though the replay assertions passed. Those are different meanings of “pass”: the fixture correctly preserves evidence that rejects the original FTS-only assumption.
SQLite's collation documentation describes how string comparison uses a selected collating sequence. The alternative fixture explicitly uses BINARY for the code field. That narrow choice is not a complete identifier policy: it does not decide whether the real product should normalize case, spaces or other Unicode characters.
The FTS failure also belongs to the tested default configuration. SQLite documents tokenizer customization, but this experiment did not compare custom token characters or a different tokenizer. Those remain possible approaches, not validated repairs. Preserve that distinction in the decision: reject the observed FTS-only assumption without claiming every full-text configuration must fail. A broader comparison would require a separate question and evidence threshold.

Close the spike with evidence and an estimate update
The completed fictional review rejects “default FTS alone supports exact product codes.” It does not reject SQLite as a whole. Sasha recommends retaining FTS as a candidate for word discovery and investigating a separate identifier route, with the intended normalization rules agreed before implementation.
Talia confirms only the teaching requirement: the three uppercase identifiers must remain distinguishable. She has not decided how lowercase input should behave. Gabe records that deployment FTS availability, code routing, index synchronization and integration review remain unresolved. No one marks the help viewer ready to ship.
The estimate update is explicit. Before the spike, the team had an unapproved two-engineer-day assumption for “add search,” based on one default query path. After the counterexample, Gabe withdraws that assumption because it omitted identifier semantics and integration work. Adding a fixed allowance to the previous number would leave its missing assumptions unresolved.
The reason is concrete. A separate code field in six in-memory records says little about loading identifiers from the application, keeping search data aligned after edits, or directing a user's input to the correct query mode. The product must also decide whether exact lookup and word discovery are separate controls. Those choices create work the earlier assumption never included; counting the fixture's lines of code would not price that work.
The revised record proposes a separate one-engineer-day investigation: verify FTS5 in the application's bundled runtime and evaluate routing between word discovery and exact-code lookup. That is a new budget request, not an approved allocation or a production delivery estimate. Talia must decide whether the additional learning is worth prioritizing. A delivery estimate will require the team to examine the remaining implementation and review work.
This is a complete spike outcome: a rejected assumption, a bounded alternative observation, an explicit unknown and a changed planning position. It does not require a successful implementation. If the business did not need exact identifiers, the same evidence could support a different decision; the requirement, not the elegance of the demo, determines what is sufficient.
Keep the exploratory code separate from delivery
Retain enough material for another engineer to reproduce the result: runtime versions, tokenizer configuration, authored records, query inputs, expected IDs, actual IDs and the failure condition. Keep the negative result beside the working cases. A screenshot showing one successful search would conceal the decisive collision.
The fixture uses parameterized SQL, but it has not undergone a security review and does not supply a production query parser. FTS expression syntax still matters inside a bound query value. The experiment also says nothing about response time at realistic scale, relevance ranking, memory, index size, concurrent use or the lifecycle of changed and deleted documents. No benchmark threshold was defined or measured.
Label the code as exploratory and track any implementation separately. Product-ready search could require requirement clarification, data preparation, query handling, tests, review and release work that this spike excluded. Copying the fixture into an application would not complete those obligations.
If the timebox ends before the evidence threshold is reached, close with an unresolved result. State what was attempted, what blocked the decision, and the smallest further investigation that might change it. The owner can then stop, choose another approach or authorize a new learning budget. Do not represent incomplete investigation as either technical impossibility or successful feasibility.
Before your next spike, write the decision and one observation that would reject the preferred approach. Put the time ceiling and evidence handoff beside them, then agree who can authorize further work. More workplace guides are available from LandOffer.ai; close with a decision record that preserves failed checks, lists unresolved implementation work and names who can authorize the next investigation.
Sources and Further Reading
- Mountain Goat Software: Agile Spikes Deliver Knowledge So Teams Can Deliver Products, checked October 27, 2026; no revision date stated in the inspected body.
- SQLite: FTS5 Extension, query syntax and Unicode61 tokenizer, checked October 27, 2026; documentation describes configurable behavior, not this application's deployment.
- SQLite: Datatypes and collating sequences, checked October 27, 2026; local equality checks explicitly use BINARY collation.


