--- name: recommendation-development description: > Builds recommendations that are actionable and no stronger than the evidence beneath them: an owner, a decision, a change within the organisation's control, and a stated basis in named findings. Separates "do this" from "decide this", labels researcher opinion as opinion, includes doing nothing and gathering more evidence as honest answers, and removes any recommendation with no traceable finding behind it. Use for "what should we do", "write the recommendations", "next steps", "turn this into actions", "the recommendations feel weak". category: 08 Insight Development ref: 08.04 tier: 1 inherits: [K2, K3, K4, K5] --- # Recommendation Development ## 1. One-line description Turns implications into recommendations that name an owner, a decision and a change the organisation can actually make, with the strength of each recommendation set by the evidence beneath it, opinion labelled as opinion, and any recommendation without a traceable finding removed rather than softened. ## 2. What this skill is used for **The research problem it solves.** Recommendations are the most consequential thing a research report produces and the most frequently overreached. They are where money moves, and they are the part of a report most likely to have travelled furthest from the evidence. Three failures dominate. The first is the orphan: a recommendation with no finding behind it, usually the researcher's or a stakeholder's prior belief, which survives review because it sounds sensible. The second is strength inflation: a moderate implication resting on one untested comparison produces the sentence "the organisation should restructure its scheduling", stated with the same certainty as a well-evidenced observation. The third is the false answer: research identifies a decision, and the report presents a decision as though it were an answer, removing a choice the organisation was entitled to make. All three are more likely in AI-drafted work, because a recommendation slot is easy to fill fluently and nothing in the resulting text signals what is missing. This skill supplies the anatomy, the strength calibration, the orphan test and the traceability requirement that catch them. **Where it sits.** The end of synthesis, immediately before reporting. It takes implications from **08.03 Implication Development** and hands recommendations, with their bases and sign-off markers, to report architecture and the executive summary. **Typical use cases.** - Writing the recommendations section of a study that will inform a funded decision. - Reviewing a draft where the recommendations do not obviously follow from the findings. - Studies where the honest answer is that the organisation faces a choice, not that it has one. - Work where the researcher must be explicit about which suggestions are evidenced and which are professional opinion. - Deciding whether the recommendation is to act, to wait, or to gather better evidence first. - Auditing an AI-generated recommendation set before it reaches a client. **Who uses it.** Research directors and consultants accountable for what a report tells an organisation to do; client-side insight leads whose recommendations will be quoted in a business case; evaluators whose recommendations carry into policy; anyone reviewing synthesis before sign-off, since per K5 §6 recommendations receive full review without exception. ## 3. When to use it - Implications are established and the report needs to say what follows in terms of action. - A decision is open, owned by someone, and the research exists to inform it. - A draft recommendation set exists and you need to know which items are evidenced and which are assumed. - The evidence is mixed or thin and you need to calibrate what can honestly be recommended on it. - The research has identified a trade-off rather than an answer, and the report must present it as one. - The right answer may be to do nothing, or to gather more evidence, and someone needs to say so credibly. - Recommendations will be read by people who will not read the findings, which is most of them. - Feasibility questions sit outside what the research can assess and need flagging rather than guessing. ## 4. When NOT to use it - **No implications or insights have been supplied.** A recommendation built directly from a finding skips both the explanation and the organisational consequence, which means nobody can check whether the action follows from anything except the researcher's judgement. Run **08.01** and **08.03** first. Where only a finding exists and the deadline does not permit the intervening work, say what is missing rather than bridging it silently. - **The decision is closed.** Where a decision has been made and is being implemented, a recommendation to reverse it is advocacy, and it must be labelled as such with the evidence and the cost of reversal named. Presenting it as an ordinary recommendation misrepresents the situation to everyone downstream. - **The action lies outside the organisation's control.** Where the change requires a regulator, a supplier, a partner or a customer to act, the recommendation as written cannot be owned. Restate it as what this organisation can do to make the outcome more likely, or record it as a constraint rather than an action. - **Feasibility, cost or capability decide the answer and cannot be assessed.** Where the research cannot tell which of two options is deliverable, the honest output is a framed decision with the evidence for each option, not a recommendation. See step 3. - **The question is which insights matter most.** Ranking and sizing are **08.05 Insight Prioritisation and Sizing**. Prioritising recommendations is part of this skill; prioritising insights is not, and doing it here builds the rank order on the actions rather than on the evidence. - **The evidence supports no recommendation.** This happens and it is a legitimate output. Where implications are real but low confidence, or where every candidate action fails the control test, the correct answer is a decision framing plus a validation proposal, not a weakened action. A report with two evidenced recommendations and a clear statement that nothing else is supportable is worth more than one with eight, of which five are opinion. - **The recommendation would be attributed to a named researcher without their review.** Per K5 §2.8 and §3.1, recommendations carry sign-off. A recommendation set produced without a named human approving it does not leave the building. - **The recommendation concerns treatment of participants, vulnerable groups, or a use of data that consent did not cover.** That is **13.05 Research Ethics and Consent Design** before it is a recommendation question, regardless of how strong the evidence is. ## 5. Required inputs **Required.** Without these the skill cannot run. If the first is absent, stop. - **At least one implication**, with the insight it follows from, the findings beneath that with their bases, its confidence, and its assumption register per **08.03**. **If no implication or insight has been supplied, do not proceed.** Say what is missing and route to 08.03 or 08.01. - **The decision the recommendation would inform**, its owner, and its timing. A recommendation with no decision attached to it is a wish, and it cannot be prioritised, sequenced or sized. - **What is within the organisation's control**, at least approximately. Without this the control test at step 5 cannot run and the recommendation set will contain items nobody can own. **Optional, and what each one adds.** - **Constraints: budget, capability, contracts, regulation, timing:** turns feasibility from a guess into a flag with a named dependency. Without them, every recommendation carries an unassessed feasibility marker, which is honest but much less useful. - **What has already been tried, and what is underway:** prevents recommending something in flight or something that failed for reasons the research now explains. The second case is valuable and is written up as such. - **The organisation's decision-making cycle:** funding rounds, planning cycles, board dates. Determines whether "sequence" means anything, and whether a recommendation can be acted on before the evidence behind it decays. - **The cost or effort profile of candidate actions, where the client can supply it:** allows honest prioritisation on effort against confidence, rather than on magnitude alone. - **Prior recommendations from earlier studies and what happened to them:** the most under-used input available. A recommendation being made for the third time is a different message, and repeating it without saying so wastes the finding. - **Stakeholder positions, from the debrief or interviews:** identifies which recommendations will be contested and therefore need their evidence most visibly attached. ## 6. Questions to ask before starting 1. **What decision is open, who owns it, and when is it taken?** Determines whether a recommendation is possible at all and what it must be timed against. *Default if unanswered:* write recommendations against the commissioning objective, state that no decision owner was identified, and mark it for researcher review per K5 §2.1. 2. **What is genuinely within this organisation's control?** Determines which candidate actions can be owned and which must be restated. *Default:* treat control as unverified, apply the control test, and flag each item where control could not be confirmed. 3. **What has already been tried here, and what happened?** Determines whether a recommendation is new, a repeat, or a repeat with an explanation attached. *Default:* ask. If unavailable, state that prior attempts are unknown. 4. **What can this organisation afford, in money, attention and disruption?** Determines whether the recommendation set is usable or aspirational. *Default:* do not assume feasibility; flag it and name the dependency, per step 10. 5. **Is the study answering a question, or framing a trade-off?** Determines the fundamental shape of the output and is the question most often skipped. *Default:* classify every candidate at step 3 and be willing to conclude that the honest output is a decision rather than an answer. 6. **Who will read only the recommendations?** Almost everyone who acts will read these and not the findings, which is why each must carry its basis and confidence with it. *Default:* write every recommendation as though it will be quoted alone, because it will be. 7. **What would make this recommendation wrong?** Determines the conditions under which it should not be followed, which belong in the output. *Default:* state the key assumption from the implication's register as the condition. ## 7. Step-by-step methodology **The position this method takes.** A recommendation is the point at which research stops describing and starts costing money. Everything that has been careful upstream can be undone in a single sentence here, because a recommendation is short, quotable, and travels without its evidence unless the evidence is fastened to it. The governing rule is one line long: **a recommendation is never stronger than the evidence beneath it**, and the practical work of this skill is applying that rule to language, not just to logic. The second rule is nearly as important and is broken more quietly: a recommendation with no traceable finding behind it is removed, not softened. Softening an orphan produces a vaguer orphan, which is harder to challenge and just as unfounded. **1. Gate on inputs.** Confirm at least one implication is present with its insight, its findings and their bases, its confidence and its assumption register. Confirm a decision, an owner and a timing. If the implication chain is missing, route back to **08.03** or **08.01** rather than reconstructing it here. *Correct result:* a numbered implication list, each with its confidence and its key assumptions visible. **2. Map the decision space before drafting anything.** List the decisions that are actually open, their owners, their dates, the options being considered, and what has already been decided. This map determines whether each candidate is a recommendation into a live decision, an input to a decision not yet framed, or a challenge to a decision already taken. The three require different handling and the third requires labelling. Writing recommendations without this map produces a set aimed at no decision in particular, which is the commonest reason a good report changes nothing. *Correct result:* a short decision map, with each open decision named and owned. **3. Classify every candidate as "do this" or "decide this".** Research frequently identifies a decision rather than an answer: it establishes what is at stake, what each option costs, and who is affected, without being able to say which option is right, because the choice turns on a value judgement or on information the study does not hold. Presenting a decision as an answer is overreach, and it is the subtlest kind, because the output looks more useful. A "decide this" output is written as: the decision, the options, what the evidence says about each, what the evidence cannot say, and who owns the choice. It is frequently the most valuable page in a report, and it is nearly always more honest than the manufactured answer it replaces. *Correct result:* every candidate classified, with the "decide this" items written in the decision format rather than as weak actions. **4. Draft each "do this" recommendation in the four-part anatomy.** Every recommendation contains: an **action** (a specific change, not a direction of travel), an **owner** (the role that can make it, not "the business"), a **decision or change** that is concrete enough to be started, and a **basis** naming the findings beneath it by reference, with bases, per K2 §5. A recommendation missing the owner cannot be actioned. One missing the basis cannot be checked. One where the action is "improve" or "focus on" or "consider" has not been written yet: it is a topic with a verb. *Correct result:* recommendations that name who, what and on what evidence, each readable alone. **5. Apply the control test.** Ask whether the named owner can actually make this change. Where the action requires a third party to act (a regulator, a supplier, a partner, a customer), the recommendation as drafted cannot be owned. Restate it as the thing this organisation can do that makes the outcome more likely, and name the dependency explicitly. This is not a downgrade, it is the accurate version, and it is what distinguishes a plan from a wish. *Correct result:* every recommendation owned by someone inside the organisation, with external dependencies named rather than assumed away. **6. Calibrate strength to evidence, item by item.** This is the core step. Take the evidence beneath the recommendation and match the strength of the language to it, using three tiers. **Strong:** multiple independent evidence streams, adequate bases, consistent across relevant subgroups, with the mechanism understood and no material contradicting evidence. Language: "adopt", "change", "stop". **Moderate:** the evidence supports the direction, but rests on fewer streams, a limited base, or an untested comparison. Language: "pilot", "test at scale", "change on a defined subset with a pre-specified measure". **Weak:** one stream, small base, untested, or an insight at hypothesis level. Language: "investigate", "instrument this so it can be measured", "run the study that would answer it". The discipline is that the same finding supports all three of these, and which one you write is determined by the evidence state, not by how much the client wants an answer. Work an explicit example (Section 8) so the calibration is visible rather than asserted. *Correct result:* each recommendation's verb chosen from the tier its evidence sits in, with the tier stated in the record. **7. Separate research-supported recommendations from researcher opinion.** Experienced researchers hold views about the category, the organisation and what usually works, and those views have value. They are not findings of this study. Where opinion is offered, it goes in a distinct, labelled place, states the basis (professional experience, category knowledge, an analogous case), and never appears in the same list as evidenced recommendations without a visible label. The test: could someone reading this in six months tell that it came from the researcher's judgement rather than the data? If not, it is mislabelled. Where a piece of opinion cannot be given a basis at all, drop it. *Correct result:* two clearly separated categories, with the opinion section short, labelled and attributed. **8. Run the orphan check, and remove rather than soften.** For every recommendation, name the findings beneath it, with bases, by reference. If no finding can be named, the recommendation is removed from the output. Not hedged, not moved to a "considerations" section, not rewritten as "the organisation may wish to consider". Softening an orphan produces a vaguer unevidenced recommendation, which is harder to challenge because there is now less of it to challenge, and it will be quoted with the same authority as the rest. This rule is absolute, and applying it visibly is one of the strongest signals of quality a report carries: a recommendations section where every item names its evidence tells a reader that the ones present were earned. Where a removed item came from a stakeholder or a workshop, say so in the working document, because its removal will be questioned. *Correct result:* every surviving recommendation names findings; removals are logged with their origin. **9. Consider the two honest non-actions.** **Do nothing** is a real recommendation and belongs in the set when the evidence supports it: where the current approach is working and the study confirms it, where the effect is real but below the threshold at which this organisation acts, where the cost of the change exceeds the benefit as far as the evidence can tell, or where a change would disturb something working for a larger group. Written properly it names what is being protected. **Gather more evidence** is the honest answer where the decision is expensive or hard to reverse and the evidence is at moderate or low confidence. It is only honest, however, when it is specific: what question, on what population, with what base, by what method, answering by when, and what decision it unblocks. An unspecified "further research is recommended" is the closing formula of a report that has run out of things to say, and it does not count. *Correct result:* both considered explicitly, and where either is recommended, specified to the standard above. **10. Flag feasibility rather than assuming it.** Cost, capability, timing, contractual position, regulatory permission and organisational appetite usually sit outside what a researcher can assess. The failure runs in both directions: assuming a recommendation is feasible and having it dismissed on contact with reality, or quietly dropping a well-evidenced recommendation because you assumed it was not feasible, which removes the organisation's chance to decide. The correct handling is to state the recommendation, name the feasibility question, name who can answer it, and say what changes if the answer is no. *Correct result:* an explicit feasibility flag per recommendation, naming the dependency and the person or function who can resolve it. **11. Say what the recommendation costs.** Where following a recommendation requires the organisation to accept a cost, state it in the recommendation, not in a later section. Costs are not only money: a slower launch, a group of customers deprioritised, a metric that will get worse before it gets better, a partner relationship strained, a capability the organisation does not have and would need to build, or a politically difficult conversation. A recommendation presented without its cost is not more persuasive, it is less trustworthy, because the reader will find the cost themselves and will then discount everything else. Naming it also protects the recommendation: an organisation that accepts a known cost in advance does not abandon the change when the cost arrives. *Correct result:* each recommendation carrying its cost in the same passage, including the costs that fall on people other than the decision-maker. **12. Prioritise and sequence.** Order the set by three considerations in this order. **Dependency:** what must happen before something else can. Sequencing errors waste more effort than prioritisation errors. **Reversibility against confidence:** a cheap, reversible action can be taken on moderate evidence; an expensive, hard-to-reverse action needs strong evidence or should be a "decide this" or a pilot. **Magnitude and reach:** how much changes, for how many. Note that this is prioritisation of actions, which is different from prioritisation of insights (**08.05**), and that a low-priority recommendation is not a weak one. *Correct result:* an ordered set, with the ordering logic written out, and any dependency chain made explicit. **13. Assign confidence and apply the cap.** A recommendation inherits the weakest link beneath it (finding, insight, implication) and adds its own organisational assumptions, per K3 §3.6 and K4 §3.5. It therefore cannot exceed the confidence of anything above it in the chain, and usually sits one level below the implication. Per K3 §5.2, every recommendation carries an explicit confidence marker, because a recommendation is an instruction to spend money or effort and the reader is entitled to know how firm the ground is. State the factor that set the ceiling. *Correct result:* a confidence level on every recommendation, with no item more confident than its implication. **14. Complete the traceability record and attach sign-off.** Per K2 §5, every recommendation appears in the traceability map with the findings it rests on named by reference, and this is one of the three rules that does not relax for lighter deliverables. Per K5 §3.1 and §2.5, every recommendation carries a `RESEARCHER SIGN-OFF REQUIRED` marker, and a named human approves the set before it leaves. Where recommendations will be read separately from the report, which they will, each one carries its basis, its confidence and its conditions with it. *Correct result:* a recommendation set in which any single item, quoted alone in a business case six months from now, still shows what it rests on and how firm that is. ## 8. Analytical framework **The chain, and where this sits.** ``` Finding → Insight → Implication → Recommendation ↑ Adds: an owner, a decision, an action within control, and organisational assumptions of its own. Therefore never more confident than anything to its left. ``` **Recommendation anatomy.** Every "do this" item contains all five. ``` ACTION The specific change. Not "improve", "focus on" or "consider" OWNER The role that can make it happen DECISION What is being changed, concretely enough to start BASIS Findings by reference, with bases, per K2 section 5 STRENGTH Strong / Moderate / Weak, matching the evidence and the verb ``` **Strength calibration: one finding, three recommendations.** *Fictional illustration, from a city transport study.* Finding: on routes where the timetabled connection window is under six minutes, off-peak boardings fell 21% against 6% elsewhere (48 routes, operator records, untested). | Evidence state | What can honestly be recommended | Verb tier | |---|---|---| | One wave, 48 routes, untested, no passenger-level evidence for the mechanism | Instrument connection integrity so it can be measured, and run the analysis that would establish whether the association holds at passenger level | **Weak.** Investigate | | Adds passenger interviews describing the mechanism, and a second period showing the same association | Pilot a connection-protected timetable on three routes, with the outcome measure pre-specified before the change | **Moderate.** Pilot | | Adds nine routes where the connection window was lengthened for unrelated reasons and boardings recovered, plus operator running data | Adopt connection integrity as a scheduling standard for off-peak services | **Strong.** Adopt | The finding is identical in all three rows. Only the evidence beneath it differs, and that is what sets the verb. A researcher who writes row three's recommendation on row one's evidence has not made a stylistic choice. **"Do this" versus "decide this".** Where the choice turns on a value judgement or on information the study does not hold, the output is a framed decision. ``` THE DECISION What must be chosen, and by whom THE OPTIONS Two or three, stated plainly EVIDENCE FOR EACH What the research says about each option, with references WHAT IT CANNOT SAY The part that is a value judgement or outside the study OWNER AND TIMING Who decides, and by when ``` **The traceability row**, per K2 §5, one per recommendation. | Ref | Recommendation | Rests on | Confidence | Feasibility flag | Cost accepted | Sign-off | |---|---|---|---|---|---|---| **The removal log.** Every candidate dropped at the orphan check, with where it came from. Kept in the working document, because removals get questioned and "it had no finding behind it" is a complete answer only if you can show which candidate you mean. ## 9. Output format **A. Recommendation record** (one per recommendation) ``` RECOMMENDATION The action, in one sentence, with the owner named TYPE Do this / Decide this / Do nothing / Gather evidence OWNER Role that can make the change BASIS Findings by reference, with bases; implication and insight refs STRENGTH Strong / Moderate / Weak, with the evidence state that set it CONFIDENCE High / Moderate / Low, with the limiting factor named WITHIN CONTROL Yes / dependency named FEASIBILITY The question the research cannot answer, and who can COST ACCEPTED What the organisation gives up, including who bears it CONDITIONS What would make this recommendation wrong SEQUENCE What must precede it, and what it unblocks SIGN-OFF RESEARCHER SIGN-OFF REQUIRED, per K5 section 3.1 ``` **B. Recommendation set, ordered** The ordered list with the ordering logic stated, dependencies shown, and each item carrying its strength and confidence markers on the face of it rather than in an appendix. **C. Decisions the research frames but does not answer** The "decide this" items, in the decision format. Required where any exist, and their absence in a study of any complexity is worth checking. **D. Researcher opinion** Short, labelled, with the basis of each view stated. Optional, and omitted entirely rather than padded. **E. Traceability map** Per K2 §5. Every recommendation, the findings beneath it, its confidence. Not optional at any deliverable size. **F. Removal log** Working document only. Candidates removed at the orphan check, with their origin. **When the evidence is thin.** The recommendation set gets shorter, not weaker. Where the evidence supports only weak-tier actions, the output is investigation and instrumentation recommendations, honestly labelled, plus the decisions the research frames. Where the evidence supports nothing, the output is section C plus a specified evidence-gathering proposal, and it says plainly that the study does not support recommendations for action. Per K4 §1, an empty recommendations slot is not a reason to write one. The number of recommendations is a property of the evidence, and a set of three that each name their findings is worth more to a decision-maker than a set of nine where they cannot tell which is which. ## 10. Quality checks Run before anything is presented. Sits on top of K4 §8, and per K5 §6 this output also receives full human review. 1. Does every recommendation name the findings beneath it, by reference, with bases? 2. Was any recommendation with no traceable finding softened rather than removed? 3. Does every recommendation name an owner who can actually make the change? 4. Is any action outside the organisation's control, without its dependency named? 5. Does the verb in each recommendation match the tier its evidence sits in? 6. Is any recommendation more confident than the implication, insight or finding above it? 7. Is every "decide this" item written as a framed decision rather than as a weak action? 8. Is researcher opinion separated, labelled, and given a stated basis? 9. Was doing nothing considered, and where recommended, is what it protects stated? 10. Where more evidence is recommended, is the question, population, base, method and timing specified? 11. Is feasibility flagged with a named dependency rather than assumed in either direction? 12. Does every recommendation that requires accepting a cost state that cost in the same passage, including who bears it? 13. Is the ordering logic written out, and are dependencies explicit? 14. Would each recommendation, quoted alone in a business case, still show its basis and confidence? 15. Does every recommendation carry a sign-off marker, and has a named human approved the set? ## 11. Common failure modes | Failure | How to recognise it | How to prevent it | |---|---|---| | **The orphan recommendation** (the signature failure) | No finding can be named beneath it; it usually predates the study | The orphan check at step 8. Remove, do not soften. Log the removal and its origin | | **Softening instead of removing** | "The organisation may wish to consider..." | A vaguer unevidenced recommendation is still unevidenced, and is harder to challenge | | **Strength inflation** | A confident verb on a single untested comparison | The three-tier calibration at step 6, applied verb by verb | | **The decision presented as an answer** | The report resolves a trade-off that turns on a value judgement | Classify at step 3. A framed decision is usually the more valuable output | | **The topic with a verb** | "Improve the onboarding experience", "focus on younger customers" | Require a specific change concrete enough for someone to start on Monday | | **The unowned recommendation** | The owner is "the business" or "the team" | Name a role. If none can act on it, apply the control test | | **Opinion in the evidence list** | A sensible view sitting unlabelled among evidenced items | Separate section, stated basis, visible label | | **Assumed feasibility** | A recommendation that dies on contact with a contract or a budget cycle | Flag the question and name who can answer it | | **Silent infeasibility** | A well-evidenced recommendation quietly dropped because the researcher assumed it was impossible | Also a failure. State it and flag it; the organisation decides | | **The hidden cost** | The recommendation is attractive because what it costs was not mentioned | Step 11. Costs travel in the same passage, including costs borne by others | | **The closing formula** | "Further research is recommended", unspecified | Specify question, population, base, method, timing, and the decision it unblocks | | **Recommendation inflation** (the signature AI failure) | Eight to twelve fluent recommendations, several restating each other at different scales | The number is set by the evidence. Merge duplicates, remove orphans, keep the set short | | **Chain amputation** | The recommendation travels into a deck without its basis or confidence | K2 §7 and step 14. Every item carries its evidence when quoted alone | | **Prioritisation by enthusiasm** | The order reflects what the room liked | Order by dependency, then reversibility against confidence, then magnitude | | **The repeat, unlabelled** | The same recommendation as two years ago, presented as new | Check prior recommendations. A repeat with an explanation of why it was not acted on is a stronger output | ## 12. AI guardrails Skill-specific only. Universal prohibitions are inherited from **K4**; traceability follows **K2 §5**; confidence language follows **K3 §4**; sign-off follows **K5 §3.1**. 1. **Never produce a recommendation that cannot name the findings beneath it.** If no finding can be named, remove it. Do not hedge it, relabel it as a consideration, or move it to another section. 2. **Never write a recommendation stronger than its evidence tier.** Adopt, pilot and investigate are not stylistic alternatives; each belongs to an evidence state, per step 6. 3. **Never present a decision as an answer.** Where the choice turns on a value judgement or on information the study does not hold, output the framed decision with its options and their evidence. 4. **Never state an owner, a cost, a timeline or a feasibility judgement that was not supplied.** Plausible organisational specifics are fabrication under K4 §2.5. 5. **Never offer researcher opinion inside the evidenced set.** Label it, give its basis, or drop it. 6. **Never write "further research is recommended" without specifying the question, the population, the base, the method, the timing and the decision it unblocks.** 7. **Never assume feasibility, and never silently drop a recommendation for assumed infeasibility.** Flag the question and name who can resolve it. 8. **Never omit the cost of a recommendation**, including costs that fall on customers, staff or a group other than the decision-maker. 9. **Never let a recommendation carry more confidence than the implication, insight or finding above it**, per K3 §3.6 and K4 §3.5. 10. **Never fill a recommendations section to a target count.** Where the evidence supports two, the output is two. 11. **Never release a recommendation set without a sign-off marker on every item**, per K5 §2.5 and §3.1. This applies regardless of deliverable size or deadline. ## 13. Best-practice principles 1. **The verb is the claim.** "Adopt" and "pilot" and "investigate" are statements about the evidence, not about tone. Choosing the verb is the single most consequential act in writing a recommendation. 2. **Remove, never soften.** An unevidenced recommendation made vaguer is harder to challenge and just as wrong, and it will be quoted with the authority of the items around it. 3. **A framed decision beats a manufactured answer.** Research that establishes what is at stake, what each option costs and who is affected has done its job, even when it cannot say which option to pick. 4. **Write every recommendation to survive being quoted alone.** It will be. Its basis, its confidence and its conditions go with it, because the findings section is not travelling. 5. **Name what it costs.** A recommendation with its cost stated is more trustworthy, better implemented and more likely to survive the moment the cost arrives. 6. **Doing nothing is a recommendation.** It requires evidence, it names what is being protected, and it is the honest answer more often than reports admit. 7. **Specify the evidence you are asking for.** "More research" without a question, a base and a decision attached is the sound of a report ending, and clients hear it that way. 8. **Sequencing errors cost more than prioritisation errors.** Getting the order wrong wastes work that was correctly chosen. 9. **Match reversibility to confidence.** Cheap and reversible can proceed on moderate evidence. Expensive and irreversible needs strong evidence, or it needs to be a pilot or a decision. 10. **Flag what you cannot assess, in both directions.** Assuming feasibility and assuming infeasibility are the same error, and the second is less visible. 11. **A repeated recommendation is a finding about the organisation.** If this is the third study to say it, say that, and ask why it has not been acted on. That question is frequently more valuable than the recommendation. 12. **Fewer, evidenced, owned.** Three recommendations that name their findings change more than nine that do not, because a decision-maker can tell which ones to trust. ## 14. Worked example *Fictional scenario, used to demonstrate method. All organisations, figures and findings below are invented.* **INPUT.** A city transport authority commissions research after off-peak bus ridership fell 14% over two years while peak ridership held. **Insight (08.01, moderate):** off-peak passengers are not choosing between bus and car for a single journey; they are deciding whether to make a chain of errands at all, and an unreliable connection makes the chain unplannable, so they abandon the trip rather than the mode. **Implications (08.03):** punctuality is measured per service rather than per connection, so the failure is invisible on the performance report; concessionary pass holders make the most multi-leg off-peak trips and are most exposed; fare-led promotion addresses a price decision passengers are not making. **PROCESS.** *Step 2, decision map.* Two decisions are open: the annual timetable specification, owned by the network planning lead, decided in eleven weeks; and the off-peak marketing budget, owned by the communications lead, decided next month. A substantial ticketing system procurement is already committed. *Step 3, classify.* The timetable candidate splits. Whether to protect connection integrity is a real trade-off: doing so lengthens some direct journeys, so it trades a benefit for multi-leg passengers against a cost to single-leg passengers on the same routes. The research establishes the mechanism, who is affected and roughly how many, and cannot say which group's journey time matters more. That is a value judgement, so it becomes a **decide this**, with options, evidence for each, and the owner named. *Step 4 to 6, drafting and calibration.* The connection-integrity measurement recommendation is drafted with owner (performance and planning lead), action (add a connection-integrity measure to the off-peak performance report, defined as the proportion of timetabled connections achieved within the published window) and basis (three findings, named, with bases). Its evidence state is strong on the measurement gap itself, which was verified against the authority's own reporting definitions, so the verb is "adopt". The scheduling change sits at moderate: the association is untested at passenger level and comes from one operator's records, so the verb is "pilot", with the outcome measure pre-specified. *Step 8, the orphan, and the judgement call.* A candidate arrived from the stakeholder workshop: "improve the journey planning app". It is plausible, it is popular internally, and nothing in the study supports it. No finding names the app, the interviews describe timetable connection failures rather than information failures, and the one app-related item in the survey shows no association with off-peak trip abandonment. The judgement was whether to retain it as a lighter suggestion. It was removed, with its origin logged, and the reasoning stated in the working document: softening it would have left an unevidenced recommendation in a list where every other item names its findings, and it would have been quoted alongside them. *Step 9 to 11, non-actions, feasibility and cost.* Fare-led promotion: the recommendation is to **stop**, and its basis is the insight plus two findings; this is a "do nothing more here" that names what it protects, which is budget being spent against a decision passengers are not making. **Gather evidence:** whether an abandoned trip chain becomes a car journey or no journey at all was not measured, and the two have very different consequences for the authority's access and carbon objectives; specified as a trip-diary study among concessionary pass holders, base of roughly 400, answering that question before the next timetable round. **Feasibility flag:** any scheduling change sits inside an operator contract whose terms the researcher has not seen; named, with the contracts lead as the person who can resolve it. **Cost:** the pilot lengthens journey times for single-leg passengers on the affected routes, stated in the same passage, with the group who bear it named. *Step 13 and 14.* Confidence: the measurement recommendation is moderate, capped by the insight above it rather than by the measurement evidence. The pilot is moderate. The stop-promotion recommendation is moderate. Nothing is high, because the insight is moderate and nothing downstream can exceed it. All four carry sign-off markers; the traceability map lists each with its findings. **OUTPUT.** Three "do this" recommendations, one framed decision, one specified evidence-gathering proposal, one removal logged, a feasibility flag naming its owner, a stated cost with the group that bears it, and a note that the researcher's view about long-term compounding effects on concessionary passengers is opinion from category experience and is labelled as such in a separate short section. ## 15. Advanced usage **Recommendations for a decision that has already been taken.** Where the evidence contradicts a committed decision, do not write an ordinary recommendation. Write the case explicitly: what was decided, what the evidence now shows, what has changed since, what reversal would cost, and what the research cannot judge. Label it as a challenge to a committed decision, and route it to a human before it reaches a client document, per K5 §2.3 and §2.5. This is one of the few places where the framing matters as much as the content. **Recommendation sets across multiple audiences.** A board, a function head and a delivery team need different levels of specificity from the same recommendation. Write one recommendation with one basis and vary the elaboration, never the substance. The failure to avoid is a board version whose verb is stronger than the delivery version's, which happens easily and quietly and means two different claims are circulating with the same reference number. **Portfolio and sequencing work.** Where a study produces several recommendations that interact, build the dependency graph explicitly before ordering: what must precede what, what shares a resource, what becomes unnecessary if something else succeeds. Sequencing on a graph rather than a list routinely halves the apparent workload, because two or three items turn out to be the same change described at different levels. **When the client wants a stronger recommendation than the evidence supports.** Per K4 §9, do not refuse without helping. Say plainly what the evidence supports, offer the strongest honest version (usually a pilot with a pre-specified measure rather than a full adoption), and name what evidence would license the stronger claim and what it would cost to get. In practice a well-specified pilot is often what the organisation needed anyway, and it can be delivered inside the same decision window. **Tracking recommendations across studies.** Keep a register of what was recommended, when, at what strength, and what happened. This turns a repeated recommendation into evidence about the organisation's decision-making, which is frequently more valuable than the recommendation itself, and it prevents the same weak-tier suggestion being reissued at strong-tier language two years later when nobody remembers the original base. ## 16. Skill chain **Recommended previous skills:** - **08.03 Implication Development.** Hands over consequences with their organisational assumption registers, which is what makes a recommendation's reasoning inspectable. This is the required predecessor. - **08.01 Finding to Insight Development.** Supplies the explained insight and its confidence, which caps everything here. - **08.05 Insight Prioritisation and Sizing.** Determines which insights are material enough to warrant recommendations, and supplies any sizing with its ranges and assumptions. - **08.02 Pattern Identification.** Establishes the scope of the structure a recommendation addresses, which determines how broadly it should be applied. **Recommended next skills:** - **12.01 Research Report Architecture.** Places recommendations where a decision-maker meets them, with their bases and confidence intact. - **11.02 Executive Summary Development.** The most-read and least-checked part of any report, per K2 §5, so every recommendation appearing there must appear in the traceability map. - **14.01 Insight Activation and Socialisation** and **14.02**, which take an owned, sequenced recommendation set into the organisation. **Runs well alongside:** - **13.01 Research Quality Review** and **13.03 AI Output Verification.** The orphan check and the strength calibration are the two tests these should apply here. - **13.04 Bias Detection**, particularly where a recommendation matches a position held before the study. - **K5**, at the sign-off this skill mandates on every item, per §2.5 and §3.1. --- A Yazi Supplied Skill and resource.